Response to NHTSA Request for Comment on AV Regulation

I know it’s been awhile since I last posted. I’ve been a bit busy with moving, a new job, and getting my second novel edited! When I saw this request for comment, however, I knew I needed to respond. Framework for Automated Driving System Safety, though because I got busy with the holidays, I missed the comments opening and closing! Luckily, they did a 60 day extension, which is where my comment was submitted and will soon be attached.

This post will be a bit denser than usual since it’s a direct response to a fairly dense framework proposal, as well as a bit longer because the framework was quite comprehensive in its construction and the analysis of implications. The gist of the ANPRM (advance notice of proposed rule making) is that the Department of Transportation has been releasing recommended rules each year for about 4 years now, and they’re considering what the next steps should be. Testing, actual regulation, more comprehensive voluntary guidelines. They lay out the elements they believe make up an automated vehicle, how they’ve been tested in the past, and existing standards like UL standards. There’s also the question of what is currently in their statutory authority, which I didn’t comment on because I’m not an expert in what DoT can and cannot do. There isn’t much more to say that isn’t in the comment itself, and the comment stands pretty well without having to read the ANPRM, so I’ll just paste it in!


From: Paul Calhoun –

February 1, 2021

Subject: Framework for Automated Driving System Safety Docket No. NHTSA-2020-0106

Docket Management Facility, M–30

U.S. Department of Transportation

West Building, Ground Floor, Room W12–140

1200 New Jersey Avenue SE Washington, DC 20590.

To Whom It May Concern:

It is with great anticipation that I see this notice of proposed rulemaking.  The need for meaningful AV regulation has grown from matter for individual States to a significant national concern.  After reading the ANPRM, I believe that responses to the questions asked at the end will encompass most of my comments, so I will begin with those.  I’m doing so out of order because I believe that my vision and concerns start from a different place than most, that of a robotics expert looking at regulation rather than a regulator looking at an autonomous system.


Question 6. Do you agree or disagree with the core elements (i.e., “sensing,” “perception,” “planning” and “control”) described in this document? Please explain why.

Question 7. Can you suggest any other core element(s) that NHTSA should consider in developing a safety framework for ADS? Please provide the basis of your suggestion.

I believe that the core elements conflate two major elements together, namely “Planning” and “Prediction.”  To take a hypothetical pipeline: Sensing perceives an object and sends data about it to Perception.  Perception classifies the object.  Prediction determines where the object will be based on its classification and both past and ongoing data from Sensing.  Planning determines how to modify the existing path plan and behaviors based on that prediction and classification.  Control executes the new plan.

I would propose that Prediction be raised to the same level of importance as the core elements laid out in the ANPRM.  Looking at the fatality of a pedestrian in Tempe, the failure was as much one of Prediction as Perception, but if you take Prediction out of it, not of Planning, Sensing, or Control.  The pedestrian’s classification changed repeatedly, and the Prediction subsystem failed to make use of existing data to predict the pedestrian’s path.  The Prediction subsystem, in fact, was configured to only use prior data for as long as the Perception subsystem maintained a single classification, so it was effectively throwing out valid data because Perception had failed.  Had the Perception subsystem settled on a single classification or the Prediction subsystem used data based only on past motion, it is likely that the vehicle would have stopped much sooner.  A good Prediction subsystem does not require Perception to make a final decision on the nature of an object to determine its likely path; it can begin that process using the object’s velocity as observed. 

A preliminary classification of solid object would have also been a significant improvement, though this does not impact the core elements.

Prediction isn’t just determining where objects have been and will be.  It’s determining where they have been and will be relative to the vehicle in motion.  That’s predicting both the motion of the object and the motion of the vehicle with respect to them, which can be a separate calculation that goes beyond using the Planning to determine future states of the vehicle.

I would add an element that is not considered here at all, that of Communication.  Communication, whether it be to other vehicles, humans around the vehicle, or infrastructure, will be a crucial part of AVs in the future.  While it is not part of the pipeline used for vehicle behavior in a vacuum or on a test course, Communication is key for both the vehicle’s overall environmental awareness and for the awareness of stakeholders regarding the vehicle’s intent.  To leave Communication out of the equation is to diminish rather than improve overall vehicle capabilities, since humans do everything from listen to the radio for road closures to gesture at pedestrians to clarify their intent at an intersection.  Even if the form of this Communication is different, the overall result must remain at least as good.


Question 1. Describe your conception of a Federal safety framework for ADS that encompasses the process and engineering measures described in this document and explain your rationale for its design.

Question 2. In consideration of optimum use of NHTSA’s resources, on which aspects of a manufacturer’s comprehensive demonstration of the safety of its ADS should the Agency place a priority and focus its monitoring and safety oversight efforts and why?

The prior answers to questions 6 and 7 reveal a way of looking at AVs which differs significantly from how automobiles have been considered until now.  AVs are vehicles, but they are also agents.  It is not enough to put them through the same type of testing as existing vehicles because those types of tests would, at best, test the Sensing and Control elements with a minimal meaningful testing occurring on Perception, Prediction, and Planning.  We must think of AVs as both vehicle and agent, car and driver. 

Let’s start with the driving exams and tests given to those who are to be licensed.  Something analogous must be administered to an AV.  The benefit is that a more comprehensive test can be given to an AV than a human because the answers given and behaviors shown by one AV applies to all AVs of the same configuration and software version.  Realistically differences in sensor performance and minute differences caused by physics will mean that testing a sample of a 5-10 AVs in a single configuration will be required.

There are several exams and tests given depending on the individual State.  I will take one of the more stringent set since they will show the way for something as important as testing an entire fleet.  First, testing the comprehension of the rules of the road, road signs, etc.  Second, the ability to perceive when those rules are to be applied (ex. eye test, hearing test).  Finally, the practical test to see how the driver is able to cope behind the wheel, first in a controlled testing course, and then on a representative section of roadway.

How does give a driving test to an AV?  We need to break down these tests and understand an important underlying assumption.  The reason why a combination of written test and test of the senses (again ex. eye test) is that humans fall into a certain cognitive space.  If we can see a road sign and we know what it means, we are assumed to be able to react in the correct manner.  The practical test checks this with a basic road sign, usually a stop sign.  This works because the majority of human brains fall into an understandable and bound cognitive space.  We think alike because our meat brains are the same layout.

An AV does not have the same cognitive space as a human, so our tests don’t work on them as well.  A Perception element is not a visual cortex; it can’t generalize nearly as well but it does function much more predictably.  The cognitive space of an AV is bounded by the hardware and mathematics which underly its algorithms.  Thankfully, the vast majority of hardware in an AV evolved from Von Neumann architecture, and the vast majority of algorithms evolved from backpropagation, Bayesian probability, etc.  It is not a perfect match, but it gives a helpful basis in understanding where to begin with testing an AV by itself.

The by itself is important, because the AV as a single unit is not the only matter being considered, but it’s the first thing we can look at because as a self-contained unit, we can think in terms of being performance based and much less about prescriptive regulation.

In terms of the (singular) core elements of Sensing, Perception, Prediction, Planning, and Control, the closer to the middle of the list, the more important the test becomes.  That is, the core elements in order of importance would be something like Perception, Prediction, Planning, Sensing, Control.  This much can be understood from accidents and fatalities so far observed, and simply because testing of sensors and control are already mature.  Any vehicle currently on the road has a well-built Control system, and testing of Sensor performance is easily done as a unit test.  Sensor Fusion perhaps less so, though that is under the umbrella of Perception, which is highest on my notional list of importance.

If we think of it from the perspective of a student driver going in for their license, we need to make sure they can see the pedestrian, know what to do when one is in the crosswalk, and guess with acceptable accuracy where they will be relative to the vehicle at any given interval.  Or perceiving a road sign, knowing how to obey it, and then determining how obeying that sign will affect future states of the vehicle. 

My highest and equal priority would be on the following things:

  1. Accurately determining a range of objects in the vehicle’s view
  2. Accurately determining those objects’ motion – if any.
  3. Accurately determining what the vehicle should do in relation to the object to maintain safety for all stakeholders.

There may or may not be a set range on any of these.  That is, a set time in which the vehicle should reach X confidence rating (confidence ratings being entirely subjective to each algorithm is a significant issue which requires a standard), a set distance from said object, or a set time/distance by which the vehicles replans safely.  This is because of the Control aspect.  Different vehicles have different stopping distances, and that should be accounted for in determining if the vehicle makes its decisions in a safe time frame.  The most important result is that the vehicle processes all data and decides in time for Control to safely execute.

However, a good result based on bad information should not be a successful test.  It must take the correct action for the correct reason, or else it is highly likely to take incorrect actions later.  One of the benefits of AVs over humans is that we can look much more critically and accurately at why the system behaved the way it did.

With all that set up, it’s time to turn outwards to the aforementioned Communication element.  It’s not enough that the vehicles can take the correct action in any given scenario, because many scenarios require communication with an external stakeholder.  Whether it’s retrieving data after a crash, emergency highway lane redirections (ex. all lanes become outbound in a hurricane), or communicating the vehicles’ plans to a pedestrian in a curb cut.  All these things require robust V2X communication in various forms.  Being able to send and receive this data will be an important part of testing going forward, and will likely be the first place where specific regulations will be required.  Performance-based testing is fine when the data involved exists only within the AV in question; how they move information around within their system is their own decision.  Once it needs to leave the boundary of the system of interest, a standard is needed.

Much like Open Mission Systems (OMS) [1] in the Department of Defense, we need an Open Driving Systems (ODS) standard for V2X communication.  This isn’t a major revelation; it’s been the subject of papers for years.  It will be DoT’s role to determine and require the minimum information that passes in these messages.  It’s not the purview of this comment to give an exhaustive list, but examples would be routine and emergency road modifications like closures, position data shared with other AVs, and the data exported for crash investigations.

There is also the legibility and predictability of the AVs, which is another branch of Communication, this time between the AV and external humans.  Legibility is understanding the intention of the AV (ex. turn indicator lit at an intersection).  Predictability is knowing how it will execute that intent (ex. knowing that the AV will remain in the same turn lane, or cross turn lanes partway through the turn because it will need to make another turn soon after).  Regulating this is an extension of existing standards and regulations on elements like the color, position, intensity, and frequency of turn indictors, and similar requirements on brake lights.  While some of this will merely be continuing to use existing communication methods (ex. turn indicators and brake lights), humans often look at the eyes, expression, and gestures of drivers to determine their intent.  An analogous method must be incorporated in AVs to maintain the same level of legibility and predictability.  Tests have been done using displays made to appear like eyes on the front of vehicles so that pedestrians know that the vehicle is aware of their presence[2].


Question 4. How would your framework assist NHTSA in engaging with ADS development in a manner that helps address safety, but without unnecessarily hampering innovation?

Question 5. How could the Agency best assess whether each manufacturer had adequately demonstrated the extent of its ADS’ ability to meet each prioritized element of safety?

By keeping singular AVs to performance-based testing modeled on driving license exams, innovation can be focused on results rather than satisfying specific numbers that will likely change based on improvements in technology.  Mandating and testing a unified V2X communication framework has minimal impact on innovation since it requires at most a set of middleware components to communicate with the proposed ODS, and at best helps AV developers focus on important safety-related data. 

The question to ask is can this AV pass a driving test?  The key part of this is to remember the difference in cognitive space, and the similarities.  Testing AVs using the exact same conditions could result in AV developers using simulations of those courses as training data; and a maxim in machine learning is never to test with training data.  Similarly, a driving instructor doesn’t test their student using the actual driving test.  It produces local solution spaces in computers, and complacency in human drivers.  Regularly changing the parameters of the course helps with this, and having multiple courses.  Test one vehicle in each AV fleet on all the courses, and change them regularly.  If they pass them all, that gives a high confidence that they will be effective in normal driving conditions at least. 

Unusual driving conditions can be tested, but will require specialized courses.  State involvement may be helpful in this.  As each State has its own DMV and driving test, so too can they test AVs for their individual State, increasing the variety of conditions tested across the fleet.  This will include not just the testing of the singular AV, but of its communication as well.  At a minimum, its communication with external humans will be tested as part of any normal driving test, but also its ability to ingest and comprehend the specific laws of the State and municipality, which must either be communicated using V2X or by looking up those rules based on the vehicle’s location in an internal database. 


Question 14. What additional research would best support the creation of a safety framework? In what sequence should the additional research be conducted and why? What tools are necessary to perform such research?

I recommend focusing research on creating a meaningful ODS, and on developing a version of the driving test which accounts for the cognitive space of AVs.  Ideally running concurrently, but realistically it’s more important to get the test working than the communication standard, which may emerge from industry over time anyway, especially as municipalities levy their own requirements about how much data the vehicles must provide and accept.

The tools required are best determined by a partnership between government, industry, academics, and the public.  There are well prepared Human-Robot Interaction (HRI) departments at universities like Carnegie Mellon’s Robotics Institute and at Georgia Tech’s Robotics Lab.  They would be well equipped to investigate how to translate a human driving test into the machine domain.  Community engagement will also assist in this, as local people can raise issues that would otherwise get lost in the race to a broad-brush solution.


Question 16. Of the administrative mechanisms described in this document, which single mechanism or combination of mechanisms would best enable the Agency to carry out its safety mission, and why?

Question 17. Which mechanisms could be implemented in the near term or are the easiest and quickest to implement, and why?

For individual AVs, starting with a mix of voluntary mechanisms and regulatory mechanisms would be best.  Keep the AV makers in the loop as the driving tests are developed, working with them to determine the best way forward in what they should include.  Elements like the list of objects which must be classified correctly (ex. pedestrian, bicyclist), which will change as AVs become more capable and evolve with the aid of the developers.  Having that guidance and framework will keep the developers moving towards a safety-conscious goal as well as contributing to that same goal.

The developers can say best what can be tested, and regulators what should be included to assure safety.  Finding common ground and collaborative development of the test structure will make sure that the AVs are safe as the technology becomes more mature and the mechanisms more regulatory.  Developed capabilities will build the runway for effective regulation.

Required reporting is crucial, even if the specifics of the data must remain confidential.  The overall results of analysis of the data should be made public and it’s in the interests of the developers to have a common agency collecting the data and publishing the results of analysis.  Being able to attain a safety ranking from DoT will spur safety innovation and increase public confidence in the reliability of the AVs that score well.

AV developer collaboration on the ODS will also be critical, and the implementation of a draft framework as regulatory requirement will also be very important.  In the same way as the testing of individual AVs will build regulatory runway as technology develops, so too will the development of the standard build runway for required use of the standard.  Something like it being voluntary for the first two years of release, and then each release makes the earliest voluntary release into a mandatory release.  That will give them time to become compliant as well as the opportunity to get ahead and shape the direction the standard takes.

Another place where research and stricter regulation will be required is in emergency disengagement.  AVs in the L2 and L3 range have a control inequality where the AV can disengage at any time, making the driver responsible for what occurs and how the vehicle behaves, but the driver cannot go the other way.  This is, of course, because the driver is responsible and theoretically more capable.  However, it can also be an easy way for the AV and the AV maker to avoid liability and responsibility for adverse effects.  The DoT regulations must require a minimum disengagement window in which a reasonably effective driver can undertake meaningful emergency maneuvers.  The AV must “take ownership” of its situation within its ODD, and either be able to handle emergency behaviors within that ODD or recognize the need for them within a time frame which allows for the human driver to take control and make a meaningful effort to avoid a collision or other adverse situation.  An AV disengaging within the window should be treated as if the AV didn’t disengage at all, and full responsibility be placed on the AV and its maker.

One of the thorniest issues is the current software development cycle.  Many companies practice sloppy DevOps and few practice DevSecOps at all.  The pipeline is based on a form of Agile which promotes releasing new versions over comprehensive testing and documentation.  This has been an issue in the past and will remain one in the future.  Model Based System Engineering (MBSE) has a major automotive component in its development, and most of the large legacy automakers continue to use it extensively.  Newer automakers and AV developers do not practice MBSE or likely any significant documentation practice at all.  While testing in simulation is a big component, developing in a model may be less so.  A cursory inspection of job offerings shows MBSE in constant demand from major Detroit automakers, but almost totally unseen among the AV developers.

Discussion of this aspect swiftly moves into the murky waters of monopoly.  Is it detrimental to innovation and good design for the same developer to produce Perception, Prediction, Planning, and Control?  It would improve both vehicle communication standards as well as interoperability, documentation, and robustness of subcomponents if different developers each specialized in one of those elements.  Mandating this in any meaningful way is well outside my understanding of the purview of DoT.  However, it is reasonable to require more robust software release strategies in the form of requiring unit test, regression test, and safety critical design.  The choice could be put before the developers: either demonstrate good DevSecOps and a robust digital thread, or every release no matter how small will require a separate and full test of the vehicle fleet.

Beyond the issues raised by the Questions in this ANPRM, there is the matter of the multiple layers of AVs currently being road tested.  My response to the questions mostly address private motor vehicles which will be on public highways, since they will make up the bulk of what will require testing and safety analysis in the near term.

Near-term and mentioned in the ANPRM are low-velocity shuttles.  I envision a minimal testing requirement for them since they will be driving a set route with minimal variation from day-to-day, and thus be lower “volatility” in terms of safety issues and need for changes.  If they can drive the route they are set to one day, they will likely drive it just as well from then on.  Testing them in situ with a few standard elements such as a simulated pedestrian, bicyclist, and other vehicles will likely suffice for long-term certification of their readiness. 

Near-term, not mentioned in the ANPRM, are a big problem.  Low velocity delivery crawlers.  Tracked, wheeled, and sometimes legged robots that carry one or more parcels the “last mile” between depots or small businesses and customers.  Most will not be on roads, which makes them an even bigger problem.  They’ll be on sidewalks and in crosswalks.  How they communicate with pedestrians and navigate these often-clogged byways without becoming a danger or a nuisance may or may not be a DoT issue, but a framework for States and municipalities to follow would be a helpful effort.  Already there has been an incident of one of these crawlers putting a wheelchair bound pedestrian (vulnerable stakeholder) in danger by blocking her access to a curb cut while she was in a crosswalk on a busy street[3].  This may seem like a local issue until one considers that the Federal government does have jurisdiction over some sidewalks on their own land, which means there will be a Federal agency that will need this guidance to regulate Federal sidewalks and land.  If nothing else, this is an ADA issue.

Finally, long-term, there’s the concern about trucking.  Not just the near-term carrying of the usual cargo by 18-wheelers, but autonomous trucks pulling hazardous materials.  Granting an algorithm a Class A CDL is a decision that will take a great deal of consideration, and we would do well to start considering it right now. 

I would like to thank you for your attention and for taking steps to assure us of a safe American rollout of AVs as they appear in our cities and towns.  No doubt many comments were sent in by organizations attempting to gain your attention to sell their product.  In that way, my comment is no different.  I am aware of the brain drain in the Federal workforce, as well as the difficulties being faced as more civil servants retire than are hired each year.  Since USAJOBS seems to have some difficulty with hiring, I thought I’d go directly to the source with this comment.  I have a deep interest in public policy and autonomous systems in general.  I have a Master’s of Science in Robotic System Development from Carnegie Mellon, and would be very interested in discussing a position within the government if DoT finds itself in need of more personnel in this area.


Paul Calhoun




Disability and Engineering

I’ve had several ideas lately for my next post, all of them somewhat more delicate topics than my usual fare.  I’m still planning on the post I was going to do – that of robots and sex – but first something that’s been rattling around in my brain since I did my piece referencing the starship delivery crawler that blocked a woman in a wheelchair from getting up onto the curb.  Starship has since released a statement pointing out that Ms. Ackerman was able to get past the robot, though as she replied, the point still stands that it impeded her progress and that others may not be so able as she was to work her way around it.

My conversations with people on the Twitter thread that this originated on got me to thinking along lines that are always dangerous for people like me.  Literal minded, technocratic-leaning engineers who are about as disabled as Arthur Dent and representative of little or no part of the population who is under served.  I got to thinking, “How does one even go about fixing this problem?”  Fixing, as in making it so that issues like this don’t occur again.

There are, after all, dozens of labs and projects working on accessibility in robotics alone.  Since many disabilities are invisible, it’s difficult to infer even from pictures of the lab staff how many of these projects have engineers with disabilities working on them.  Certainly, there are disabled people testing the robots and autonomous systems in order to improve their function.  The thing is that what these projects all have in common is that they are for disabled people rather than for everyone including people with disabilities.

There are foundations which help with this.  The best way to foster inclusivity in design is to have an inclusive design process where the designers represent a wide array of people who might come in contact with the system.  According to the CDC, over a quarter of US citizens are disabled.  What their statistics also show is that disability is a spectrum, and even the CDC graphic simplifies this greatly.  Each of those categories is a spectrum, which further breaks things down into smaller and more specific types and severities.

This is important because, to make a crass numeric argument, some disabilities make it difficult in greater quantities for the person to become an engineer, and they all create a bar of some kind.  10% of undergraduate STEM degrees are awarded to people with disabilities, 6% graduate of any kind, and 2% PhD.  This means that most labs are extremely unlikely to include any kind of disabled postdoc or faculty, and startups aren’t faring much better.  After all, the average startup has only 5 employees, which means that – with the greatest regret from the founder, no doubt – there’s a strong chance that they will see hiring a disabled engineer as expensive, and as long as they’re below 15 employees, they’ll use the excuse that the ADA doesn’t apply to them.

It also doesn’t help that almost half of all disabled people are over the age of 40 (again, let me stress that I’m using US stats because they’re available and I’m a US Citizen.  Your mileage may vary in other countries).  This matters because the average age of the founder of a tech startup is 45.  They tend to attract recent graduates who are more risk tolerant.  That’s still 50% of disabled Americans, but that’s 50% of somewhere between 2 and 10% of all engineers, depending on what level of education the startup needs.  Startups also skew towards graduates, so let’s say 5% of 50%, or 2.5% of engineers are in the startup range and have a disability.  Factoring in the likelihood that this startup is in a trendy post-industrial wasteland building – the sort of big concrete and steel cube that was built before modern accessibility concerns – and there’s a huge barrier to entry.  Not to mention the aforesaid spectrum.

So, let’s take it as read that right now there’s no way to get a representative sample of disabled engineers on a design team.  And lest anyone point this out, I’m aware that they’re not the only people who are unlikely to be represented on a design team.  What’s the solution?  In general, the only one that I can think of is the obvious – the solution given to any under-representation – “hire and train more disabled engineers.”  There are certainly foundations and a plethora of articles suggesting just that.  The other answer – the one that will “have to suffice” until we get more people trained – is mindfulness and inclusivity using methods like collaborative design, a process which has shown outstanding results in designing for neighborhoods.  This, of course, puts the onus on the disabled community to come and explain their needs, but to be fair the process does that to everyone. It’s just that with fewer disabled people and so many types of disability, someone may be the only person in their neighborhood who can give their point of view accurately, and not everyone has the time to become a full-time advocate to help tech companies do the right thing in the community.

As I said above, I’m not representative at all.  I’m sure I’ve missed plenty and that I may even look and sound like yet another engineer who needs the situation explained.  I’m sure it’s tiring to have to keep educating people.  Speaking of education, maybe the whole issue needs to be flipped.  In addition to more disabled people being educated in engineering, more engineers need to be educated in disability.  In mine, for example, accessibility was treated like a specific application rather than a pervasive and constant design element.  We designed to help people specifically with disabilities instead of designing to help everyone.

I’m reminded of a competition we had where engineering and business students had to work together to pitch to a mock-entrepreneurship panel made up of business professors.  The topic was to design something to help the elderly.  See how we’re already segmenting?  Not help everyone including, just help them.  There’s the classes we help specifically (disabilities, elderly, youth, autistic, etc) and everyone else.  Never the twain shall meet or else things get complicated.

It also amazes me sometimes when a team goes into the meeting and the lead says, “we need ideas about how we can help X person” and they get all surprised when I reply, “have you tried asking them how we can help?”

This was more rambling than I usually do, but I expected it.  I’m venturing into deep waters without much awareness of where I’m going.  Which is probably why so many engineers feel lost at sea when they think about this topic.  And yes, I’m keenly aware and somewhat embarrassed that I didn’t try – you know – just asking someone in the community for help with this article.  But then again, would I be properly representing a clueless engineering team if I did something as sensible as that?

Godbots, AV Bills, and Rage Against the Delivery Drones

It’s been a few weeks since my last post, but as I get close to a 50-minute talk at Philcon, I think I should take some time away from promoting my new book (gotta mention it once a post!) and check on how things have been going in some of the topics I’ve written about before.

First, a bit on the robots in religion. My #1 reader (hi mom!) sent me this article and I pulled the thread a little to see what the newest developments are. The main article deals with Mindar, an anthropoid robot that recites a 25-minute pre-programmed sermon.

What leaps out at me in the monk’s words is that he hopes the robot itself becomes a quasi-deific item. That as the monks come, live their lives, and die, an AI that will one day be loaded into Mindar’s body will continue to learn and grow wiser. That the robot will eventually achieve enlightenment or channel the goddess of mercy as an animate, intelligent incarnation. This has so many possibilities; humans as builders of gods. Robots as links to the divine. Robots as religious relics, tended by the monks and priests.

There have also been a couple more robots that take care of ‘rote’ worship. In India, there’s a robot arm that performs a ritual motion.

Elsewhere, a device is being used to recite the name of Buddha or says prayers endlessly, in the manner of a prayer wheel only purely digitally. While neither intelligent nor a classical robot, it shows the march of progress.



On the autonomous vehicle front, the AV START Act and SELF DRIVE Act in the Senate and House respectively have been abandoned. SELF DRIVE passed a couple years ago, but the AV START Act never did, and its abandonment leaves a gap. The House and Senate formed a joint committee in July, planning for industry involvement and cooperation between the two legislatures in drafting a new autonomous vehicle bill. This has so far gone nowhere. Having nothing else they can do, DOT has sent around $60 million in grants to institutions in seven States, a mixture of university research groups and State Departments of Transportation.

While a fairly predictable development, the impeachment investigations and almost inevitable Articles and trial will leave the House and Senate unlikely to work together on anything this year, and nothing that isn’t a crisis until 2021. Meanwhile, autonomous shuttles are being rolled out nationwide, from Kalamazoo to the Brooklyn Navy Yard. All while Waymo partners with firms in France to do a 20 mile mixed highway and city street shuttle from De Gaulle to La Defense and Tesla continues to make grandiose claims about “Full Self Driving” capability (which is apparently not what it says on the tin because they still have the “always keep your hand on wheel and eye on road” disclaimer) which includes existing “Smart Summon.” Let’s have a look at Smart Summon in action.

And a longer video of the Tesla hitting the garage wall

But that’s just features. We’re here for policy. Well, let’s see what the police think about policing non-existent AV policy.

6:00 mark if it doesn’t take you right there.

And we hear it! The line we’ll get from every driver of every ADAS car from now until a law is passed. “I wasn’t driving it!” And the cop saying what most will say when confronted with no regulatory guidance, “Then who do I write the ticket to?”

It’s a classic question, but one that is going to be asked with increasing frequency. When an autonomous system does wrong, who is liable? Ticketing or arresting sleeping drivers in Teslas on Autopilot is clear enough; there’s an operator asleep at the wheel. Whatever the features involved, that’s a well established situation. When there is no one in the car at all, we hit the issues in this video. In this case, I’m surprised the owner didn’t get a citation for operating the vehicle in a dangerous manner – or whatever ticket you get for trying to do a stunt where you steer from outside the car or let go of the wheel and run alongside.

It’s also worth noting that it’s hard to be sure, but the car didn’t appear to stop because there was a police car behind it, but because the owner stopped it or it reached its destination. This has happened before. Police had difficulty getting an autonomous Tesla to stop when a driver was passed out at the wheel.

Let’s look at the Smart Summon disclaimer

Smart Summon is designed to allow your car to drive to you or a location of your choosing, maneuvering around and stopping for objects as necessary. Like Summon, Smart Summon is only intended for use in private parking lots and driveways. You are still responsible for your car and must monitor it and its surroundings at all times and be within your line of sight because it may not detect all obstacles. Be especially careful around quick moving people, bicycles and cars.

In other words, “if this screws up, it’s your fault for not paying attention” just like Autopilot. The question on my mind, then, is how fast you can stop the car if you see it about to hit something. Given the damage so far, probably not fast enough.



Our final update is on autonomous delivery drones. This was not the story I expected to see as the catalyst for a “torch and pitchfork” moment, and that’s the exact lack of awareness that has caused the issue in the first place.

In a post from Emily Ackerman, a disabled PhD Student at the University of Pittsburgh, we hear a report of a Starship (previously talked about in the Autonomous Delivery Crawlers: A Policy Perspective article) robot keeping Ms. Ackerman from getting out of a crosswalk by blocking the pavement cut. This turns out not to be a bug but a feature.

It makes sense from the perspective of an engineering team that likely had few if any differently mobile members. The robot needs to cross the street. It should communicate this in the most legible way possible. So it places itself in the curb cut to signal its intent. It shouldn’t cross when there is an obstruction, so it waits for a clear path and a ‘walk’ signal if possible. A person in a wheelchair would likely be wide enough to stop it from proceeding, and so it gets itself stuck right where the person in the wheelchair needs to go in order to get out of the street. Starships are probably not programmed to back up, and seemingly don’t have many or any sensors back there. The robot is pretty well stuck by its weight and the high traction surface, so it can’t be moved.



Image result for starship robot back

Starship Robot from the back (source)

What makes this more ridiculous is the following video

It can climb the curb. Probably decreases its mean time between failures by stressing the motors and joints, but it’s built to do it. It doesn’t even need to descend the cut to get into or out of the street. Placing itself there helps communication, but there are other ways to signal that.

As an issue, it’s a reminder that what seems like safe behavior for a robot may taking up space intended for the safety of human traffic. What is truly outstanding are the replies!

I wouldn’t have the slightest problem kicking it out of the way

i will personally fight this robot

Let me know if you need me to come by and give it the old Philly special.

That comment is referring to the destruction of a hitchhiking robot in Philadelphia

carry a screwdriver around and scrap em for parts

And aluminum baseball bat and/or a sledgehammer should solve this issue.

Gonna smash the shit out of it if I ever see one

I’m cherrypicking, but that’s a pretty good number of violent responses. No doubt (at least mostly) non-serious, but the fact remains that this is a serious issue that caused a great deal of public anger. Starship pulled the robots off the street when they got the tweet, so they recognized it as a hazard that needed to be addressed immediately.

The problem will grow, though. These are the first of many, and already crowded sidewalks will be choked with them. Perhaps they’ll migrate to the streets and bike lanes, becoming closer in form to humans on ebikes, scooters, and motorcycles, but the fact remains that they will take up space, and often dangerously. The proliferation of delivery people is a hazard I know well from living in New York, where many behaved quite dangerously, risking themselves and others.

The shift over to robots will focus that annoyance that was previously spread across many individuals onto single companies. After all, if one robot behaves badly, that means they all will in the same situation. A couple more incidents like this, and the industry will die. Some may say rightly. It’s up to the developers to prove the detractors right or wrong by designing the delivery crawlers to be either safe or unsafe, but as we see in policy everywhere, it only takes one bad actor to get an entire group thrown out.

These issues are developing daily, and I hope to get back to Pittsburgh or another high tech city soon so I can experience them firsthand. If there’s any takeaway here, it’s that some companies (Starship) learn from their mistakes and others (Tesla) just keep going ahead hoping for the best. “Move fast and break things” is an OK motto with purely digital, non safety items, but when the tech is embodied and safety critical, what you break might be a human body.



Longer Mindar video:

The Soul of a Robot Artist

The We Robot conference I recently attended had several outstanding presentations and papers on issues facing AI/robot ethics and law.  You can see an archived livestream where you may find the author asking questions (quite frequently, in fact).  This article deals with what I believe were my first and last questions at the conference: what makes an AI author/artist different from a human one, and does a machine spirit confer the ability to make art?  Let me be very clear from the outset that the two concepts I’m about to combine in this paper are well outside my usual bailiwick.  My spiritual knowledge has been demonstrated by my Robot God article (anyone who has read it can come to their own conclusion).  My knowledge of copyright is mostly that I want one, and some basics of its application.  My understanding of art – inferring from the behavior of an art professor I worked with – is somewhere in the negative region.  The metaphysics and social concepts of what makes art art and why some things are art and other things aren’t art is something I freely admit confounds me to the degree that I do worse than chance when trying to separate what is from what isn’t art.  In fact, I somewhat suspect that my result would be less than 10% on the art / not art test.

With that admission of being worse than ignorant on one of the three concepts I’m about to attempt to talk about, let’s get into a discussion in which I will try my darndest to at least be accurate even if I’m not right.

The first was brought up in the workshop day, so there is no paper to go with it.  This was the idea of Shinto with respect to robots.  As an animistic spiritual structure, the question was what role the spirit of inanimate objects plays in AI and robotics.  I am already wrong, however.  A brief look at Google and it’s Wikipedia entry tells me that Shinto isn’t strictly animistic anymore.  Since this premise was being discussed by a predominantly Japanese and highly educated panel, I can at least be mistaken in good company.  The discussion we had suggested that some limited quantity of classically inanimate objects and forces have a spiritual component, and that a robot or AI may be one of those.  Let’s move on with the discussion with the assumption that they can.  Even if there is no major animist religion I can name just now, the possibility that robots have spirits remains.

This has many implications, but let’s stick to art for now since that’s the premise of the article.  Time to be even more wrong than I was about Shinto!

The copyright panel (I’m linking to We Robot’s program because all the papers were preliminary and so it isn’t good form to directly cite) concluded the conference on a fascinating and contentious topic: can an AI produce copyrightable art?  Let us take it as read one more time and then not bring it up again: I am not an expert on copyright, and I know even less about what a romantic author is.  However, I think I got the gist of the argument against AI authorship. The central points seemed to be:

  1. An AI author’s work has no intrinsic artistic value because it is the algorithmic combination of existing works.
  2. An AI cannot be an author because authorship must change the author.  I used the word “improve” in my question, which appears not the be the best term.  I think the idea is that no matter what the impact on the consumer, if the author has no concept of the universe of thought which makes up human civilization, does not contribute to this universe, and is not in some way altered by the act of authorship, they aren’t worthy of protection or the appellation of “author.”
  3. An AI has nothing to gain by being protected as an author.  They will create regardless of whether they are protected, they have no use for money, and recognition is immaterial to them.  Therefore anything that might be of value culturally should be in the public domain so that all may use the output immediately.

I as a technocratic philistine disagree to some extent with all 3 points (although I sincerely hope that I have restated them with moderate accuracy).

First, it’s important to distinguish an AI author from an author who uses AI to augment their capabilities.  There’s no easy place to draw the line, but I think that the example of Harry Potter and the Portrait of what Looked Like a Large Pile of Ash works in many ways for this article.  First, it was made using a predictive text generator which I have tried out in a graphical format that the website provides.  You have to specify whether you are doing dialogue or text, as well as ending a sentence when you want.  You can “seed” the sentence with a word or phrase to start, and pick from a list of 18 suggested words.  I think this still makes it an AI author since the input can be quite minimal, and contributing structure still leaves the content to the AI.  The “voice” belongs to the training set data.

This brings me to point 1.  The voice of the author is derived from the training set. Therefore the act of writing – should the content created be added to the set – can only reinforce what has been written.  That is, the voice never changes because what comes out cannot never evolve and is a sort of closed conceptual structure.  Sort of like if you build a machine that shines a light which is on the other side of the color wheel from a light shone on it, and then you reflect the output back to the input (it just goes red -> green -> red -> green, and cannot escape this loop).  This would be a perfectly valid argument to me if the AI was made with a training set already picked out and no new inputs other than what it had created (and perhaps not even that).

The thing is, why not have an AI that looks further?  If it were to consume material at varying intervals, with a bias towards the genre it’s supposed to write, it wouldn’t be that much different in its reading habits than I am.  That probes the boundary between replication and influence.  Even with the Harry Potter AI, it created a plot which JK Rowling is almost certain not to have ever conceived of, and did so in her authorial voice, making it arguably some of the best fanfiction out there.

Going on to the similar point 2: An AI author is not changed by the act of authorship, has no intention or meaning that they themselves understand as such, and thus cannot create art.  Let’s be clear about one thing, we’re talking AIs in the near and possibly mid term future. Far future with sentient/sapient machines, totally different ball of wax.  I dislike but understand this argument.  I think it brings in too much ambiguity to be enforceable, because it assumes that humans are ipso facto artists by their very nature.  That a toddler putting together random paint is changed by the act – and possibly they are.  It also assumes that the mind and soul of the artist and author are the core of protectable art.  It doesn’t matter what we as viewers and readers experience, it’s the creator that is important.

This is crucial, because it means that the Harry Potter story referenced above is not art or at least not protectable art.  The fact I found it funny, insightful, and a clever reflection of how an author’s voice in print is as distinct as it is when heard, regardless of whether it makes sense, is immaterial.  The fact that it spawned a myriad of dramatic readings and animations online all of which are protectable is immaterial.  It isn’t protectable because we don’t believe it changed the author in a way we consider artistic.  It may have altered the text predictor in an algorithmic way, but because we can mathematically (eventually) derive the precise change, it isn’t spiritually/intellectually/culturally/whatever valuable.  Of course, I suppose the flipside of that is that it can’t be infringement. Does that break the chain?  Is all the derived work of a piece that is axiomatically in the public domain thus not infringement as well, or does this public domain for AIs exist in a sort of artistic subspace where the chain simply skips one inspiration and links the derived works to Harry Potter itself?

Here’s a question, if mechanistic predictability and understanding is the quotient here, what happens when the human brain is completely mapped?  What if we find that we can mathematically model our own minds?  Do we then cease being artists because we can create a function for our brains that transforms input to output without needing us?  A computer capable of simulating a human brain would likely be able to simulate something more complex.  Do we lose our ability to be artists when something that can treat us the way we treat a modern AI comes along, ceding art to the most mystically unintelligible intelligence?

See, that’s the thing for me. The following image is artistically unintelligible to me:

As the young people say, An Art

I trust that someone could eventually explain its artistic merit, message, and contribution to human culture.  Trust, that’s the key word there.  We believe that we are being enriched by paintings of soup, or color splashes, or geometric shapes arranged in a certain way.  It doesn’t need to tell us a story because we believe it represents a story to the person who made it and that story meant something. Barbara Cartland may eventually have a bibliography of nine hundred books.  Imagine how changed she must have been by the time she died!  Or perhaps she started approaching the asymptote of authorial alteration, a literary Zeno’s paradox.  What’s the Planck length of artistic and cultural evolution, the atomic quantity of personal growth?

We trust each other as human creators to create something that at the very least means something to us.  If we don’t extend that trust to AIs, we’d better have a reason that doesn’t apply to our own creations as well.  We assume that there is change occurring in the human creator, but what if the robot does have a soul?  Does that change the equation?  Don’t we then have to give it the same consideration?

With Ms. Cartland we reach point 3.  She needed protection not, perhaps, as a meritorious creator but as a person who needed to make a living.  An AI is not motivated by money, survival, or anything but the programmatic imperative to create.  I won’t argue the point directly because I don’t (directly) disagree.  We protect IP partly to secure the rights of creators and encourage people to create more and reap the benefits.  I’ll just make one yikes factor point: isn’t it the goal of most artists, at least to start out with, to create simply because they want to?  To be able to throw off the limitations of what is profitable and just make something because they want to?

So I guess that I agree with the conclusion if not the premise.  An AI doesn’t need protection because an AI has, by its nature, achieved the end goal of an artist: it is free to make what it will, without fear of starvation or penury, and without the demands of patrons and customers.  An AI isn’t undeserving of a copyright because it’s lesser than we, but because it’s transcended the need.

Getting back to the spiritual side, if robots do have souls, and the soul of an artist lives in an AI that makes art, it’s a heck of a lot closer to the ideal than most human artists.  And isn’t that the biggest kick in the teeth for us as biomechanoids: we may hold the intellectual high ground, but our spirits aren’t looking quite so shiny.  If a war robot’s spirit is harmed by being a war robot (and maybe it isn’t if the souls of warriors are improved by engaging in combat), then an art robot’s soul is commensurately improved by its unhindered artistry.  If that’s the case, it’s goodbye to points 1 and 2 because there’s the improvement right there, in a realm where mathematics can’t be used to deny the robots the status as authors.  As for point 3, I suppose we could start charging them for electricity and processing power.

I might be getting a bit artistic with my arguments in the last couple paragraphs, but hey, I’m allowed.  I’m human.

Robot Tax Policy Part 1 – Before

I recently had the pleasure of seeing Brianna Wu on a couple of panels about AI. One of the policies she was proposing was “taxing the robot which takes your job.” I asked if she would be open for an interview, which she was, and hopefully I’ll have that for you in the near future. Before I went back to Ms. Wu, I wanted to dive into the issue myself with the currently available materials so that I could present a before and after of my thoughts on the idea. Here’s the ‘before.’ I’ll go through the literature and write down how I feel after reading the articles and papers I find on the topic.

Let’s start with the problem and a (moderately) dense paper on the topic. Acemoglu and Restrepo present a very well researched and defined background on automation from the industrial revolution to the present and how – so far – automation has produced only temporary disruption to the overall job market. The main issue facing any disruptive automation is task-based: what tasks are automated, what tasks remain, and what tasks are created as a result of the automation. If the sum of tasks gained is positive, then the disruption is likely temporary. If the sum of tasks is negative, then the automation is probably eating away at jobs. The complexity of these tasks is also taken into consideration, with more complex tasks being weighted more heavily in the equation as being better for overall quality of life (under the assumption that they pay better).

Take AI as an example. An example of a low-complexity task created by AI is the necessity of feeding examples into a classifier. AIs can learn to classify data (images being the most prevalent), but require enormous amounts of curated data to function. Thus, there is a need for humans to find pictures and label them as either containing what the classifier is looking for, or not. However, this job is so low-complexity that the CAPTCHA people are using it now. People who want access to websites find pictures containing cars, which proves they’re human, and those pictures are then passed to an AI training set. However, a high-complexity task created by AI is setting them up to run, supervising them, and improving the algorithms.

Another key takeaway is the concept of excessive automation. This is a perverse incentive in the tax code which makes it cheaper to automate than to have a human employee even in cases where the automatic process is less productive. Another example: a machine that costs $1,000,000 with a 10-year useful life, and a human who is paid that in the course of 10 years in salary and benefits. Because of the option of accelerated depreciation in many tax codes, the company can write off 50% in the first year, and then a steadily decreasing percentage over the remaining 9. In the first 3 years, the company may write off 900,000 of that million dollars. Meanwhile, the employee is not only not a depreciable resource, but actually costs the company in payroll and medicare taxes over and above their salary and benefits. Thus, the machine may produce 30% less than the employee, but is still a better investment. True, there are maintenance costs to consider, but there’s also the possibility it will last much longer than the expected 10 years.

This is bad for the government as well. In addition to losing out on the revenue from the written off depreciation, there is also the loss of the worker’s income tax, and the payroll/medicare taxes. In the current structure, automation is heavily subsidized. South Korea recently debated robot taxation, settling so far on lowering the amount of write-off they allow for automation equipment.

In the next paper by Abbot and Bogenschneider, the authors start with a brief introduction to the issues around automation, concluding with the assertion that the current industrial revolution is fundamentally different from previous ones. They follow with an overview of the taxation problems with automation. An interesting fact that is not included in the previous paper is that while it may seem as if a sales tax or VAT would help with this, they do in fact make the problem of automation worse. This is because a machine doesn’t buy goods and so don’t incur the end-result costs of VAT. Thus, though a company may buy more and thus pay more VAT with an automated process, they then send that cost down the line to the end consumer, which is usually a human.

After that, they get to the main part of the paper where they compare different tax strategies to try to curb excessive automation. I will present them in brief along with my take on what difficulties they would face:

  • Reduce or eliminate the deductions around automated processes. These include the capital depreciation previously mentioned, as well as business expenses around provisioning and repairing the machine. The authors further suggest that the more automated the company, the less they are able to deduct. This is the simplest solution and likely would remove several of the tax incentives for automation.

Problems: Mainly those of categorization. What is an expense and what is an automated process? Is it just the direct costs around a piece of equipment, and what types of equipment are considered automation for the purposes of the tax code? If a company, for example, buys a robot that uses the same raw materials to produce goods as the worker it replaced, are the raw materials now included or are they allowed because they would have been purchased whether or not there was a robot? If the robot is faster and more material is purchased, would we tax the difference?

  • A “layoff tax”. Some States already track layoffs and charge more in unemployment insurance to firms that have greater layoffs over the course of the year. This would be expanded to the federal level and possibly include a way to determine whether the layoffs were motivated by decrease in corporate size or by increases in automation, perhaps by looking at year-over-year sales.

Problems: Categorization again, as well as fears of offshoring caused by the increase in the tax rate. Also, how long do we levy this tax? Is it based on layoffs for one year, five, ten, or what? Can a company hire more people and thus lower the rate, and if so, what if they’re just hiring low wage bodies to decrease their tax burden after laying off a lot of higher wage people?

  • A “human worker” tax incentive. This can either be a depreciation analogue for future wage costs, or repealing the payroll/medicare taxes.

Problems: Decreased tax revenue for the government either way. In the first case, it would significantly decrease revenue across the board. In the second, this would also effectively end Medicare, Social Security, and Unemployment Insurance.

  • Tax automation indirectly by looking at productivity per capita, as a ratio of either profit:employment or sales:employment.

Problems: Is there an average ratio they have to exceed before the tax is levied or is this a down-the-line to everyone ratio? What is the tax rate and how does the ratio affect it? That is, is it 1% for every 100:1 profit/person over a certain limit, or something more nuanced? The authors themselves suggest that sales ratio is a bad idea since it would disproportionately affect low-margin, high-volume companies. I think that this proposal may be the most fair, but would also be extremely hard to pass legislatively. Companies would successfully lobby to have the ratio set so high that it would tax almost no one, and even if in later years automation got to the point where the ratio was passed by most companies, they’d lobby to have the ratio increased based on spurious reasons.

  • Related and from the previous paper. A tax on that ratio, maintaining payroll and medicare taxes at the same quantity even after layoffs, or charging a tax on robot deployment.

Problems: The same as previous, plus these: If the layoffs weren’t automation-based, it punishes companies that do layoffs to try to survive in a competitive environment. If there’s a tax directly on robotic deployment, it brings up the categorization issue again, and is a tax on an entire industry without a view of whether tasks, wages, and employment are actually increased by some robotic deployments.

  • Direct increase to the corporate tax rate

Problems: Not only assumes that automation is an across-the-board problem that all firms should pay for, it also makes companies more likely to automate so as to get deductions from depreciation. Also from a practical standpoint, this is a nonstarter. Our government is unlikely to raise taxes on anyone for any reason ever. Arcane rule changes are one thing, but a straight up tax hike would die before it even reached a committee.

Out of these, I think that decreasing the deductions is most likely to pass (not requiring much predictiation power since it already has in SK, though likely not happening during this administration), but that ultimately the ratio of employees to profit would have the most societal benefit. Some would say it’s taxing business success because the owners have found ways to make more money with less outlay. To that, I say, yes. Yes it is. But you can’t really call yourself a “job creator” when you then say that success is the elimination of jobs.

Let’s look at the media discussion and policy movement so far. Starting with Bill Gates’ remarks which really set the ball rolling on taxation of robots and automation. He didn’t say what method he favors, leaving it instead to policy people to figure it out. Gates’ main thrust is that a worker can produce X amount, and a robot can produce the same, but the human worker is taxed, while the robot isn’t. This type of inequality is nothing new, and one might even call it a core facet of capitalism. The business owner owns the means of production and so they reap the benefits.

Lawrence Summers, former treasury secretary and economy adviser to Presidents Clinton/Bush and Obama respectively, rebutted this in (for a civil servant) very strong language. His arguments, in brief, were first that there is no way to draw the line, second that it improves living standards, and third that taxing robots decreases output and thus decreases overall wealth. The first argument has been covered in this article previously: it is difficult if not impossible to legally define “bad automation”, and even harder for an organization like the IRS to administer it. There will be either inevitable collateral harm, or a law so weak that almost no one will pay the tax. The third argument presupposes that robots and humans are currently on equal footing in the tax code, which has been discussed previously as well.

His second argument, that the goods and services provided by automation are naturally better than those of a human, is questionable. His examples are autonomous vehicles, online booking, and robotic surgical tools. The first does not yet exist, the second is arguably only efficient from the perspective of the services being booked, and the third isn’t (yet) replacing any labor. To take the second and third in more detail: An online booking system as a replacement for a travel agent is attractive to consumers because it means they spend their own time rather than spending money for a solution. This may take more time (depending on what the person needs) or less time (because they just wanted something quick that a search algorithm can easily replicate), and replaces expertise and judgment with algorithmic probability. I wouldn’t call this better, merely different. The third example increases the safety of surgery and decreases patient recovery time. This is good for just about everyone, and because the demand for surgery is so great, there is no significant loss anywhere. Should the robot get to the point where they can autonomously do a heart operation in an hour and send the patient home in two days, the impact would be far greater.

Let us now cherry-pick a counter-example: automated ordering kiosks at fast food outlets. This removes a job and lowers costs for the business, which will likely pocket that difference, write off the cost of the kiosk in depreciation, and thus screw the tax collector coming and going. There is a minimal change in the efficiency of ordering food, and because of the lack of expertise in using the system, the customer takes longer to order, requiring more kiosks than there were workers, which are all written off. There is an increase in kiosk-making employment, which would be an offset except the kiosks are likely made and assembled by robots.

Later that year, governments considered – and usually rejected – any kind of automation tax. The EU refused to do so, citing fears it would slow innovation. It was put forward in San Francisco, but it appears that so far only a tax on automated vehicle rides (not yet implemented) has been approved. The right-wing jabbernowls briefly took notice of it via Ms. Wu (who’s notoriety in other spheres makes her a focal point for the attention of extreme 1st amendment writers). They made no policy arguments, but the reaction was illustrative of how the debate would likely be presented in the Senate if the congress-critters believed the average voter was watching or cared.

In South Korea, however, the ball began to roll. As the most mechanized economy in the world began to notice that automation was beginning to become economically inefficient, they debated, proposed, and finally passed a measure which decreased the subsidies for automation. This was met with a mixed response, but it was a first step and given that SK remains a robotics hub globally 2 years later, it seems not to have stifled innovation.

Talk has continued in the US in 2018 and 2019, with little change in arguments for and against. The only concrete action that has been taken is the most recent GOP-sponsored tax bill, which made it possible to not only accelerate depreciation, but write off the entire cost of automated equipment the first year it is bought. Effectively an even greater incentive to automate, there are already signs that it has resulted in more automation, though the results remain preliminary. This is, however, to be expected. Some States and smaller countries may take action sooner, but the larger governments will not act until we have passed the crisis point, and the actions they take then will have to be extreme to correct the imbalance. Until then, pro-business lobbies will make sure that automation is the preferred solution, speeding up the job loss to a rate where new tasks cannot organically be discovered fast enough. We may not be in the big one, the moment when automation starts taking tasks away faster than it adds them, but the tax code and pro-automation policies currently in place will make the short-term situation feel like it.

Ultimately, the difficulty is that as the means of production focus ever more on machines, it is harder for we as humans to link ourselves to what is produced in our economy. If fewer people produce more goods, how do we as a society represent that output? Money is a rough representation of the value of time spent doing a productive task, and is exchanged for other peoples’ output. What is the value of money when everything we need as human beings is made by an autonomous system without any input from a human except as the owner of the system? What can we give them that is more valuable than what they already have? This is why universal basic income is gradually being taken more seriously; because we can’t be sure that this isn’t the big one. A post-scarcity society is in an even more Malthusian bind than a scarcity-based economy because its output is strictly measurable, mechanistic, and changes only with the increase in technological advancement. There is no “hidden hand” to absolve society of blame, no faceless economic force to excuse poverty and starvation.

And that may be the most dangerous thing of all. When everything is automated, who chooses the size of the population, its distribution, its upkeep? As long as economics is confusing, policy can be vague. When we know precisely how much our machines can produce, we can say with certainty that X people can live at Y standard or X+N people can live at Y-R standard. We can optimize the curve. Then it’s just a matter of determining how to control X in order to get our desired Y.

Robot God

The phrase ‘Deus Ex Machina’ is thrown around a lot these days, but it’s been a few millennia since gods were regularly delivered by mechanical devices (here in Western Civilization, anyway. There’s a strong argument that some Eastern religions have been doing it all the time, discussed later).

This article was kicked off by me remembering my visit to the Human Robot Interaction conference in Chicago last year. There was a scholar by the name of Gabriele Trovato there and he was presenting his paper on Theomorphic Robots. Like many papers presented at conferences, this a guideline rather than a research paper; referencing existing efforts and attempting to make some rules about it. I admit it’s a bit of a juicy topic, but if people are going to start building prayer robots, I think perhaps some guidelines are in order. A very helpful article about Professor Trovato’s work along with an interview is at IEEE.

The guidelines are straightforward and sensible, mostly to do with respect and keeping the mystical aspect of the divine by having the robot do as little moving as possible, and to reduce as much as possible user input to tell the robot what to do. A rule and ideal that seem to contradict is that the robot shouldn’t ‘impersonate’ the divine, but also that having these robots in the context of religion helps people be more comfortable with a functional robot doing some task (like keeping track of the user’s health). This, to me, seems like trying to have it both ways. The robot is being subtly manipulative by using religion as a mask (his word), but then shouldn’t be mistaken for true divinity but merely a tool/intermediary. This combines with the rule ‘don’t call it a robot’, which seems even more in conflict with the previous rule.

As part of his research, Professor Trovato has in the past produced two robots. One is the SanTO, a small robotic saint in a niche which keeps the user company and may – if the church allows it – one day help teach catechism. The other is a Daruma doll which assists its user in health tasks. Both do a task which other, secular, robots also do. My understanding of Daruma dolls is very minimal, though a quick search tells me that there are some traditions where they are burned at the end of the year. Images of the famous Core explosions of Aperture Science spring to mind.

Statues of Saints are something I feel I have a stronger grasp of, which is what concerns me. My knowledge of Catholic practice is basic, but I believe that people pray to their figurine Saints as channels for the Saints themselves. That is, you pray to the image of Saint Jude, for example, and the Saint in Heaven hears it. I think that this is why the icons are supposed to be 100% inanimate. People pray to the icon, and if the Saint chooses to answer, they do so in whatever manner they deem appropriate. Having their simulacrum bless someone could be a trifle presumptuous (I will avoid the B and H words, lacking the expertise to be sure of accurately using them).

Design Strategies for Representing the Divine in Robots

Photo: Gabriele Trovato/Waseda University

Theomorphic robots may, paradoxically, be on firmer footing when they’re performing genuine religious activities. A funeral business in Japan is offering a Pepper robot to recite the ritual prayers, and has the option of livestreaming as well as recitation and ritual motion. The article brings something up which occurred to me as soon as I started thinking hard about these robots, which is the tradition of the Prayer Wheel/Mill in Tibetan Buddhism. These automated prayer objects imbue good energy and remove bad energy when turned by the faithful, and are often considered just as effective when powered by heat, wind, water, and electricity. The engineer in me desperately wants to know if the rpm affects how much holiness is generated.


The BlessU-2 is a similar idea. It will dispense a blessing in seven languages and two voices. If you want, you can print the blessing out and take it with you, in case you get audited by the Blessing Revenue Service and need to provide receipts for all the blessings you’ve gotten that year. The Church that commissioned it reportedly has no plans to replace priests with the robot, but considering how blessing congregations is a rote task, perhaps they’ll reconsider.

A more ambitious project is the Xian’er, the Buddhist Monk robot in China. It appears to be using a degree of AI chatbot architecture to provide Buddhist advice to people who ask it questions. Still in the early stages, updates are on the way, and it’s the only robot on this list that is allowed to provide advice by the religious authorities that govern it.

I wasn’t sure what to think after Professor Trovato’s presentation, but I think I’ve formed my opinion. I disagree with him, and I’m almost diametrically opposed to his proposed rules. I don’t think that we should cloak utility in holiness. A machine is a machine, and a machine in the theological sphere should assist in religious duties which a machine may perform. I included the alternate translation of “Prayer Wheel” as “Prayer Mill”, which is more accurate, and which shows that at least one religion has a place for machine in god. A Prayer Mill is unashamedly mechanical, and does not replace the need for spiritual effort on the part of those who benefit from it. Instead, like robots automating repetitive tasks, it frees the human spirit to contemplate something greater while the machine spins in the background, doing the repetitive but necessary task of purifying air, water, heat, and electricity.

To place a robot in the theological space but not call it one, and to have it’s use and value be in a secular (if worthy) frame of helping someone do a daily, but not holy, task is fundamentally dishonest to me. Many religious organizations have a secular side that helps people, like feeding them and running homes for the elderly or disabled. This is part of the ministry, but is as much about the spirit of the provider as the giver.

One does not pray to the wheat thresher nor the mill which contribute to the flour for the communion wafer, nor to the filtration system through which water flows before it reaches the pipe which delivers it to the font. We do not pray to the industrial machinery which shapes, bakes, and stamps communion wafers. Textile mills which produce vestments do not demand faith and praise.

Not yet, anyway.

Praise the Machine God. Deus Est Machina!

Copyright Games Workshop

Cute Robots

Robots are going to get cuter. It’s a natural defense strategy when dealing with the public. As I discussed in my article on robots being bullied, there are going to be a lot of people who will see robots out in the street and be impolite or hostile towards them. Have a quick look at this delivery robot from Starship (best part is at 1:40 if it doesn’t start there for you).

The robot moves in a straight line, asks politely for people to get out of the way if they block it, and an alarm goes off if it’s interfered with. Starship says they’ve had good results from people on the street helping the robot if it gets into trouble. It may be a lasting effect, or novelty factor. The thing is that there’s work being done to help people live with their robots. This clip from Star Wars shows us a (remote controlled) small wheeled robot that’s a little more lively than the Starship delivery crawler.

It’s cute. Not just because it’s small, but because it squeaks and blats in a way that tells you what it’s thinking. It’s trundling along making a happy noise, and then encounters an angry Wookie, at which point it squeals and runs away. That behavior is more effective – in my opinion – than a polite ‘please move’ or an alarm.

The Starship robot is trying to act like a peer, and is relying on etiquette and politeness, which only works if the robot can enforce its existence as a social equal of the person it’s asking to move. Since AI is nowhere near sentience, and most people know that a pizza delivery droid isn’t going to be the first robot to get a genuine person brain, there are going to be humans who take this as a challenge to their own personhood*. What I believe is more effective is the robot acting like a trained animal or a pet. Not using words, but instead physical motion and expressive noises to make itself seem less intimidating and less like it’s trying to take a social position that may cause aggressive responses. It not only disarms it to people who are predisposed to be hostile, but also engages bystanders in feeling sorry for it when mistreated.

Speaking of motion, there’s something else at work in that mouse droid’s actions. It’s weaving around and making turns in a fluid, organic manner that helps show what it’s thinking and where it’s going. In robotics, this is called legibility and predictability. I’ll borrow one of their graphics to explain (there are better ones in the paper, but this one is compact):

The gray line is predictable. If you know the robot wants the right-hand bottle, it’s the path you expect it to take. Direct, and goal-oriented. The orange line is legible. If you don’t know what bottle the robot wants, it tells you what bottle it will take. Indirect, but communicates more information. Like the mouse droid above, it tells you what it’s doing as part of the motion. The droid probably could have spun its wheels in the opposite direction and been gone without turning around, but by doing that little turn, it tells us that it will be going away now and communicates its distress in the little tremble of its turn. The weaving as it approaches our heroes also seems to be telling them that it wants their attention. It’s deliberately curving around to get in front of them, rather than taking a straight-line path near the wall or across the floor.

Land Rover is taking an interesting tact on these kinds of nonverbal cues:

One of the big issues with a driverless car (or any robot with perception) is the lack of a human giving human body cues. A driverless car probably sees you, but there’s no indication. This car not only has the lights telling people whether it’s planning on stopping or continuing, but also the eyes like a human driver aimed at a pedestrian to tell them it’s aware of their presence. The eyes serve a secondary function in sweeping back and forth to help people think of the cars as watching the road. The car communicates both awareness in general and awareness of specific things in a way that humans around it can quickly and easily understand.

There’s also a real ‘Brave Little Toaster’ vibe happening here.

“I’m just a car/vacuum cleaner, doing my job. Just another day in the office/meadow…” The car says with its eyes and the straight line below that looks like a mouth. The impression of boredom makes the car seem commonplace and workaday. “Nothing novel about me,” the car says “there’s a million of us all doing our jobs every day everywhere.”

With robots, like with humans, the way the message is sent is as important as what the message contains. Words are one of many channels in human communication, and often considered to convey less meaning than some of the others. When robots are sending impressions with a glance, a twitch of the servo, and the way they move their joints, we’ll be much closer to having them as much a part of our daily lives as service animals and strangers on the street.

* Parallels in human history and interaction occur, but perhaps are best left to the more tactful scholars of such sensitive issues.

The Autopilot Defense, AV Ethics, and Why Neither Work

A couple of days ago, an article from Elektrek appeared in my feed about a man from the Netherlands who tried to argue that he shouldn’t have been issued two tickets because it was his Tesla Autopilot that was driving at the time. The judge ruled that the level of supervision required to use a Tesla in Autopilot mode meant that the driver was responsible for the car’s behavior. Much like how the fatalities in the US (some of which have been covered here) are deemed driver error, so are traffic violations. This is good for Tesla’s liability, and if the stories get enough coverage, maybe the threat of being ticketed for something the Autopilot system did will encourage some more responsible behavior.

Also in the last month or two I’ve been hearing more about AV ethics. The Moral Machine project at MIT set up a series of scenarios and the tools to make new scenarios, and millions of people judged what the AV should do. The premise was similar to the Trolley Problem.  The AV detects its brakes are gone. It can sense the number and type of passengers as well as what is immediately ahead of it (barriers, people, animals). The people in the car or on the road can be old, young, rich, poor, fit, male, or female in various combinations. There are two lanes and the question is whether to swerve into the other lane.

An illustrative example from the Moral Machine

They then compared attitudes across multiple behavioral ideas between the countries they got enough responses from to be significant.

The differences between the US and China with the worldwide average

The ethics problem is an important one, and it’s good that we are looking at it now so that when we’re ready to implement that kind of judgment, we’ll know what is socially acceptable behavior in an emergency, and how that behavior should be adjusted based on locality. However, the publicity this has garnered (it was even a segment on this season’s QI) has had the effect of perpetuating the belief that AVs are more advanced than they are. People see these studies and assume that if it’s being studied, the cars must be in a state where these decisions can be made. It’s a natural assumption. The cars can see, therefore the cars can see what is there.

My contacts in autonomous vehicle design and my knowledge of computer vision tell me otherwise. The only human that the AV is absolutely sure exists in the one inside the car. Everything else is a confidence figure based on an imperfect feature matching classifier (or some other variant of computer vision algorithm).

Let’s have a look at some footage from the output of a Tesla Autopilot. I use Tesla because it’s out there now, prevalent, and there’s a decent amount of video that can be used for examples. Any AV will look similar but not the same, especially if it has a LIDAR array (which Tesla doesn’t) which adds a point cloud to the camera and radar returns. Disclaimer: I am applying general knowledge of robotics and autonomy to a limited data source. I know nothing of the specifics of Tesla’s internal logic, and my views are inferences based on knowledge and observation that I put together to give you the most likely explanation.

Let’s start with a clip from a video taken on the streets of Paris. We can start anywhere because the big thing I want to talk about is everywhere in the video.

The boxes. Color coded red for vehicles, yellow pedestrians, green bikes, and so on. First word is what it is, then the estimated distance. Then below that in parenthesis is the kicker: the confidence factor. For vehicles the factor is usually high or becomes high very quickly. Pedestrians, not so much. Their boxes fluctuate more, and often the vehicle identifies distant inanimate objects as pedestrians – which is fine because it’s cautious and at that distance probably not a big deal at street speeds. However, real pedestrians often have fluctuating boxes that come and go, and their confidence factors can start out very low and never go up. Also, try 1:47 where a piece of building has a pedestrian confidence rating of 100%. Caution is good, but that’s a very high confidence for something completely wrong at 55 meters.

Real pedestrians are often ID’d and never go above 50% when the car is much closer. Starting at 10:05, there’s a group that crosses right in front of the car and it takes over a second for their boxes to be consistently yellow and confidence factors above 50%. Would you want your car to run into a barrier because it was 30% sure there were between 2 and 5 people ahead of it, and that 1-3 of them might be a motorcycle?

I don’t say this to condemn Tesla, I say it to make it clear that the ethics problem cannot be considered real-world one right now. The premise requires a degree of technical sophistication that is not yet there. We may have it in the near future, but I strongly doubt that even vehicles as advanced as the Waymo cars can maintain a consistent view with high confidence factors – and they have LIDAR arrays to add another sensor layer.  The addition of thermal cameras would likely help this even more, though at the cost of slowing the algorithm which has to fuse yet another sensor type.  Surety is pointless if the car can’t “think” in real time.

Based on this degree of uncertainty in what the vehicle can perceive, the current priority that I heard murmured regretfully in the halls of the Robotics Institute is this: do everything to save the passenger. When given a choice where only one option may harm a passenger, the AV will always choose the other option. Therefore in the barrier case above, the AV of today doesn’t care if there is one or five people in the car or if the entire other lane is filled with babies. It will swerve. It will do so because it knows it will save a human life that way.

It may sound bleak, heartless, even immoral, but it’s sound reasoning.  Having an AV kill a pedestrian to save a passenger may be a tragedy, but it’s worse if it kills a passenger to save a bundle of tumbleweed that climbs just over the confidence threshold during the fraction of a second it is making its decision in.



Bonus footage: Here’s some Tesla view from highways. I think it’s fun to see how the gore problem (getting trapped in crosshatched zones which can eventually result in hitting a median) has been coming along.

Also, something fun to look out for in all of these videos is the label on the bottom half of the confidence box. The top as I noted before is ID, type, and confidence rating. The bottom is what lane it believes the object to be in relative to the car, it’s speed (also relative to the car, that’s why so many are negative; that means the car is catching up to them), and the radar return from them. The radar return is the bottom text, which says either ‘moving’, ‘stopped’ or ‘stationary’ or ‘no rad sig.’ Looks like it’s using Doppler shift from the radar arrays to figure out the motion of objects nearby. Interesting that there’s a ‘stopped’ and ‘stationary’ rating. That sounds like they’ve got some prediction going where the Autopilot is looking more closely at objects it believe will move at some point ‘stopped’ rather than ones which will not move at all ‘stationary.’

The first video was taken around September and I think is part of the same Paris series. Line colors appear to be red = can cross line, purple = don’t cross line, yellow = unsure. Let’s see how it’s doing.

I started at 3:52 and in that instance, the gore is almost perfectly captured. There is a very brief period of red/yellow where the car might merge into it, though without access to the algorithms, I couldn’t say whether it would do it or not. There’s another, smaller, gore at 4:07 that is equally well handled, and again at 7:52. There are a few more as the video goes on, but the most important is 13:14, where the car is briefly headed directly at a gore. It labels it purple but crosses anyway.

I’m going out on a limb here and saying that the merge was at the behest of the driver, and to be fair it isn’t on the purple for long. That’s the danger zone if someone looks away and lets the AP handle it, though. It might keep merging right, but also might refuse to cross the purple, stay in the gore, and hit the median. To be honest, I might have considered that a very unlikely edge case except for the incidents I’ve previously covered, one of which was fatal.

The second, captured and uploaded sometime in the last month. Watch the center screen. It looks like there are different color codes here. Green = can cross line, red = don’t cross line, yellow = unsure.

The gore ( at 3:27 if it doesn’t start there for you), is yellow at first, but turns red just after the car gets to it but before the crosshatch is reached.

There’s been a definite improvement. In fact, it might be the case that AP has overcome the gore problem completely. On the other hand, all of those lanes were very clearly marked. Apparently Paris has found more resilient paint than Silicon Valley. The most recent video I can find of a car aiming itself at a median is footage I’ve used before, which was from April. Promising. I hope they’ve figured it out.

Autonomous Vehicles and Net Neutrality: Life in the Paid Lane

Net neutrality pundits on both sides have likened the internet to a highway. Mostly – up until now – it’s been like a single lane on a public highway or road. Whatever is on it can only go as fast as the car in front of it, and the only way to change that is to change lanes or go on a different road entirely. In a few places, you can even pay to be in a faster lane – though this isn’t always effective. What you can’t do is pay to speed up the lane you’re in, or do what ISPs are supposedly going to do with the end of net neutrality, and send some traffic through the pipes faster than others. Not yet.

The idea is simple enough. I get in an AV and pay more to get where I’m going faster. Not only does the car go at a higher speed, but traffic routing software makes sure that other cars wait at intersections for me to go through first, get over into slower lanes so I can pass, and the routing software even moves cars onto other streets to clear it ahead of me. The municipal system effectively treats me like an ambulance or police car; I’m going where I’m going and nothing besides a real emergency vehicle can stop me. This is the far end of the slider, but it’s not hard to imagine a situation where there’s a demand and traffic based price for arriving x minutes earlier than the standard. Most ride shares now have price tiers based on whether the riders are willing to walk further or wait longer for a cheaper ride together, etc.

The good news for people who are thinking that this is taking pay-for-performance too far is that it may be too difficult to set up as a marketplace. Extending the pricing in the opposite direction by offering an express service mediated by vehicle-to-vehicle communication is dependent on that communication and on an infrastructure that makes it possible. The communication protocols are likely to be implemented since a communication standard will be necessary to make sure that emergency vehicles are recognized and that extraordinary situations like having every lane in a highway made outbound during a natural disaster. The problem isn’t the communication itself, but the marketplace.

Let’s say I’m Uber and I want my AVs to be able to have an ‘express’ option where they go faster and have priority in traffic. In this future, most if not all cars are AVs and move in an optimized manner to get everyone where they’re going as quickly as possible. If all cars are moving in a close-to-optimal manner (a truly optimal path usually requires more computing time and power than it’s worth), then having a car go to ‘express mode’ will cause traffic to stop being optimal in order to improve the performance of a single vehicle. Since this costs the system – and the people in the cars – something, then it will need to be paid for. Otherwise every car would try to move in a way that advantages it the most and traffic control would be impossible. On a more down-to-earth business level, the scenario goes like this: My car says, ‘I want to get somewhere sooner, so cars 2 through N have to make way for me and get somewhere later.’ The other cars reply, ‘What’s in it for me?’ If Uber wants to get their cars somewhere at the expense of Waymo, or if a Tesla needs a line of GMs to take the slow lane for a few seconds, they’ll have to pay for that privilege.

That’s where it all falls apart. What incentive does a carmaker have in participating in the Expedited Arrivals market? Sure, they might make money from people wanting to go faster, but then they have to pay the others for it. The margins are likely to be razor thin at best, or else the price for the consumer would be very high. Imagine picking up an AV at JFK airport in New York and wanting a fast trip to your hotel on Atlantic Avenue, or coming into Reagan in DC and taking an AV cab to Capitol Hill. That cab would need to negotiate with and find a reasonable price with thousands of cars. It can all be automated, but it adds up. A second to you has to be worth much more than a second to them or else they’d win the bid and you wouldn’t be able to pass because not only do you need to pay them to make way, but you also need to pay the city for the right to go faster, possibly extend lights, and then there’s the fee that the AV company would take for negotiating the deal.

On the other hand, I say as I look at this and start to think it might be possible, perhaps the economics work if you modify your expectations. Let’s say I tell my app where I want to go and an AV arrives. Before and during the trip, I have options for guaranteed or variable arrival times, like hotels have for floating or guaranteed rooms. In the hotel, I know I’ll get a room on the lowest rate, but not what kind. With the AV, I know I’ll get where I’m going but not how long it will take. The more I pay, the more my AV can bid to pay cars and the city to get me there faster. If I get caught in traffic, I put an extra five bucks in the slot and the car starts bidding again and pulls ahead.

So far I’ve said how I as a passenger in an AV use my money to get there faster, but what about the people in the other cars? What happens to the money I’ve paid them? Well, that’s where the big fight will likely be. It will naturally go to whoever owns/operates the car. You pay for other people to get out of the way, but those people don’t get paid, they just don’t lose money. Instead, the payments go to the company.

If they own their own car, they might get the funds instead, though. Perhaps that’s how the concept of centralized ownership of vehicles will fail. People will realize that if they’re not picky about how fast they get somewhere, they can pick up a few bucks every trip from other passengers who are in a hurry. Imagine the marketplace where people just toodle around the block near Times Square getting paid to get out of the way instead of getting a job! Get one of those Volvos with the beds, figure out the wi-fi password for the Hotel Edison and just let the money flow in!

Pardon me while I just go and patent something…

Autonomous Delivery Crawlers: A Policy Perspective

Much has been written about the possibilities of full sized autonomous vehicles (AVs), including by me. There’s a lot of material covering delivery drones or small unmanned aerial vehicles (sUAVs) – which I haven’t covered here but have spoken about in presentations to engineers at CMU. The thing is that most packages, takeout, or other deliverable items are much too small to waste an entire car, and a delivery drone is very limited by its energy requirements and capacity. AVs are better off for big things and sUAVs for rural and suburban deliveries where the place delivering to or the place the delivery is coming from are relatively distant from a main road.

Let’s say you’re ordering takeout from a restaurant that’s within 5 miles of easily traversed road or sidewalk, or getting a package and the depot is within 10 miles of where you live. An AV is an extravagance and if everyone used sUAVs, the air would be full of them. Additionally, if there were large buildings around and the sUAV traffic is limited to flying between them, the airborne traffic jams could get severe. If you’re using a delivery robot, it’s going to be a crawler. I use the term not because they’re all on treads – most of them are wheeled – but because of the relative speeds and location. It’s a good word to differentiate autonomous delivery robots that are smaller than a car and travel on land from their airborne and human-occupiable counterparts. They’re little critters that run around with pizzas and tacos inside, ferrying small packages from one place to another. Some have a range of 20+ miles and are quite large like the Robby 2 and others are shorter haul robots that can fit a couple pizzas like the Starship.

Images courtesy of Robby and Starship

Starship Technologies has been especially quick off the mark when it comes to regulation. By July of 2017, Starship had succeeded in getting five States to pass legislation allowing delivery crawlers to operate on sidewalks. Not only that, but they took things a step further – getting the legislation to include a maximum weight limit that froze out one of their competitors. Usually an industry has to be much more mature before companies start weaponizing regulation to stop competitors, though it may be a coincidence in this case. Then again, it may not. Starship seems leaps and bounds ahead of anyone else in the media and in the testing sphere, despite not being the first company to test on public sidewalks.

On the flipside, San Francisco almost completely banned crawlers, relegating them to industrial districts for tests only and only going 3 mph on sidewalks at least six feet wide. The regulation also required operators to be within 30 feet of the robot, which all companies already did. This appeared to be spearheaded by Norman Yee, supervisor for SF’s 7th District. Supervisor Yee has had clashes with tech companies in the past and expressed concerns about traffic, job loss, and wealth inequality in regards to autonomous machines. This didn’t stop Starship, which pivoted in SF from transporting from businesses to private homes and offices to doing deliveries in office parks. This includes traveling unsupervised on the 4.3 acre Intuit campus in Mountain View where the robot delivers food and office supplies. Starship had good reason to be unruffled by the ban in one city given how welcoming other States had been, and their reception in DC where they’d been testing since 2017. The delivery robot industry recently got legislation passed in DC to permanently allow them to operate.

In fact, ‘unruffled’ is Starship’s attitude towards everything in the public space. While other companies tend towards quietly testing and making deals with service providers to deliver their food, Starship’s leadership enjoys showing how chill they are about everything. When reports surfaced about people kicking the robot, they responded that if that was how people needed to handle their anger management then fine. The robot was tougher than a foot and if things got really rough, its cameras, siren, and GPS meant that no one was getting away with beating up the robot and not paying for it.

Though San Francisco’s reaction was precipitous, it may not have been unjustified. The public have shown very positive attitudes towards delivery robots, especially ones that assist human carriers. If we use helper robots instead of fully autonomous delivery crawlers, then that will add to sidewalk congestion for sure. If we go with mostly crawlers, there will still be an uptick in sidewalk usage since many delivery services use bicycles, and there will likely be a general increase in the use of delivery services since they will probably be cheaper and more convenient. It’s a pretty clear case of the Tragedy of the Commons. Absent a clear cut reason not to – like a law or obvious liability issues – the sidewalk is taken up by more people and robots over time and eventually becomes difficult to navigate. At the moment, it seems very likely that humans will win this fight wherever it crops up because sidewalks were obviously meant for pedestrian traffic, and small slow moving robots won’t get priority over that.

The question is: where do the robots go? Like the shared electric scooters now proliferating in cities, delivery robots often go too fast for pedestrian paths (or take up too much sidewalk space), and too slow for car lanes or even bike lanes. Bike lanes themselves have had a very rocky climb towards prevalence in cities, remaining sparse, dangerous, and often nonexistent in many municipalities. The author has lived in two suburbs that didn’t even have sidewalks. How, then, can robots expect to get their own artery when human piloted vehicles are still waiting for their bike and scooter paths?

The answer may lie with another set of robots. The specter of autonomous vehicles looms in almost any discussion of future transit, and here they are again. We’ll assume that roads won’t get wider, nor will people stop walking on sidewalks. Then that means reallocating street space.

The promise – and threat – of AVs is that they will reduce or entirely remove the need for on-street parking since they can drop people off and then go on to their next task or a remote parking location. Let’s assume this is true. In that case all the on street parking will be up for grabs. I expect that the immediate demand will be for bike lanes, but with the loss of revenue, cities may have other ideas. They might require AVs to report how much they drive and then pay a toll based on their city street usage. However, that might not work out to be enough. Cities might also put in a toll for small delivery robots and let them have the onstreet parking lane instead of cyclists and scooters. It’s not likely, but it’s possible in some places. Maybe AVs will somehow reduce traffic so much that multiple lanes will be opened up for bikes and robots. A lot of things can happen.

The truth is that there are many possible scenarios, but no likely ones. The Boring Company wants to put cars on skates underground and skate them from garages to hyperloop arteries. I was going to observe that since robots don’t need air, sunlight, or as much overhead space, they’d go well in the tunnels, which would also help protect them and their cargo from the weather. However, it seems like Elon Musk wants to put us in the tunnels and the robots up on street level. Personally, I think that sounds like we’re training the robots to seek a dystopian future where they get to have parks and we’re stuck living underground, but that’s just me.

I remember a brief span after wheeled bookbags first came out. Everyone started wheeling their bags. Then it turned out that this put way too much extra stuff on the sidewalk, annoyed people, and generally got in the way so much that everyone went back to backpacks on backs. Sidewalks are congested in a lot of places, and unless/until they become quite inexpensive, these robots are going to have trouble operating where businesses can realistically buy one. Giving them their own space when bikes and scooters can’t get it in a lot of places will only increase friction, and the hypothetical space savings from AVs will probably go to bikes and parkland rather than convenience droids. They make a lot of sense for indoor use, but I think the cities – like rural areas – belong to the sUAV. In well run suburbs, there might be a cost-effective use case. Low enough foot traffic in a highly centralized market district could mean delivery robot utilization makes sense. I lived in Bloomington, Indiana for a few years, and given the quality of the food on the main street and the relatively compact residential section with well laid out, underutilized sidewalks, a handful of delivery robots shared by the various restaurants might do well. Not expecting to see them out in Odon or New York except as novelties, though. They’d run out of sidewalk in Odon and be run off the sidewalk in New York. Though New Yorkers might look at the robots that are moving consistently with traffic, and then at the tourists blocking people’s path and decide they like the robot better.