F. Piekniewski, “Deep learning has been at the forefront of the so called AI revolution for quite a few years now, and many people had believed that it is the silver bullet that will take us to the world of wonders of technological singularity (general AI). …We have now mid 2018 and things have changed. ..By far the biggest blow into deep learning fame is the domain of self driving vehicles ..
But by far the biggest prick punching through the AI bubble was the accident in which Uber self driving car killed a pedestrian in Arizona. From the preliminary report by the NTSB we can read some astonishing statements:…
Aside from general system design failure apparent in this report, it is striking that the system spent long seconds trying to decide what exactly is sees in front (whether that be a pedestrian, bike, vehicle or whatever else) rather than making the only logical decision in these circumstances, which was to make sure not to hit it. …
In fact if there is anything at all we learned from the outburst of deep learning, is that (10k+ dimensional) image space has plenty enough spurious patterns in it, that they actually generalize across many images and make the impression like our classifiers actually understand what they are seeing. Nothing could be further from the truth, as admitted even by the top researchers who are heavily invested in this field….
the problem is that the input space is incredibly high dimensional, while the action space is very low dimensional. Hence the “amount” of “label” (readout) is extremely small compared to the amount of information coming in…” Read more Hmmmm…. Very interesting. We still have an awful lot to do. See also,G. Marcus, below. Alain
F. Fishkin, May 31, “Artificial Intelligence may be able to drive better than humans most of the time….but is that good enough? Join Princeton University’s Alain Kornhauser and Co-host Fred Fishkin for Episode 41 of the Smart Driving Cars Podcast. More on the latest from Uber, Tesla and Nuro. Listen and subscribe.“
Hmmmm…. Now you can just say “Alexa, play the Smart Driving Cars podcast!” . Ditto with Siri, and GooglePlay. Alain
Real information every week. Lively discussions with the people who are shaping the future of SmartDrivingCars. Want to become a sustaining sponsor and help us grow the SmartDrivingCars newsletter and podcast? Contact Alain Kornhauser at firstname.lastname@example.org! Alain
M. Sena. June, 2018, “… The Summit was an action-packed, two-day event held on the Princeton University campus. I have tried to hit the highlights, of which there were many. The third annual Summit is already being planned by the organizers, so mark your calendar for the week of May 13th.
In Musings on the last page, I reflect on all those ‘free’ services that we receive from companies that definitely are not listed as non-profits. Do we realize that we indenture ourselves to companies like Facebook and Alphabet/Google when we click on their Accept button, or that we are transferring personal data to Alphabet when we use free public Wi-Fi? There is no free lunch, folks. I have continued my look at high-speed rail from the May issue. …” Read more Hmmmm…. Another very interesting edition. Michael, thank you for the kind review. Alain
G. Marcus, Jan 2, “Although deep learning has historical roots going back decades, neither the term “deep learning” nor the approach was popular just over five years ago, when the field was reignited by papers such as Krizhevsky, Sutskever and Hinton’s now classic (2012) deep network model of Imagenet. What has the field discovered in the five subsequent years? Against a background of considerable progress in areas such as speech recognition, image recognition, and game playing, and considerable enthusiasm in the popular press, I present ten concerns for deep learning, and suggest that deep learning must be supplemented by other techniques if we are to reach artificial general intelligence….” Read more Hmmmm…. Very interesting. DeepLearning is a statistical approach and as such provides answers that are not perfect (not that anything is). Unfortunately, the imperfections are simply wrong and the difference between right and wrong, in some instances, is substantial. This should come as no surprise simply because human intelligence is also imperfect and at times simply wrong. The implications can be substantial. That is all about being human, so in a sense it is comforting that AI may have some human qualities. Unfortunately, the same courtesies that may be offered to a human for being human may not be offered to a DeepLearning…. More importantly, if we are to surrender control to an AI entity our expectations of that AI may well be extremely demanding. So much so, that we are not anywhere near there, yet. Our appraisal of the AI needs to be less on its ability to win some competition or to get things right and much more on “why did it not discern an object ahead sufficiently well such that it would have activated the brakes and not hit it.” i.e. focus on the losses, not the wins. To be of any use, the losses have to approach zero, else we’ll turn them off, cut their wires, unplug them, just as with Hal in 2001. Alain
T. Lee, May 30, “In the last couple of years, companies like Uber, Waymo, and GM’s Cruise have been testing more and more self-driving vehicles on public roads. Yet important details about those tests have been kept secret.
Two Democratic senators are determined to change that. Last Friday, they sent out letters to 26 car and technology companies seeking details about their testing activities—part of a broader investigation into the safety of driverless vehicles…” Read more Hmmmm…. Yup. See my comments to the letter in the next section Alain
E. Markey & R. Blumenthal, May 25, “We write to inquire about your company’s safety protocol and practices while testing autonomous vehicles (AVs) on public roads….
1. a. Why has your company chosen to est there. Hmmmm…. Answers should be interesting but likely not substantive.
b. How did your company determine your AV technology was safe enough to operate on public roads? possibly interesting responses, but likely more about the role of the safety driver rather than substantive details about their closed-course testing and simulation environment(s).
… 4. … how many times have drivers had to unexpectedly regain control… not really relevant unless it also asked about the details of the encounter that triggered each of these “disengagements”. The follow-up questions sort of get to the issues, but each disengagement involves a log of data that is used by these companies to find the root cause(s) of the non-routine disengagements and to fix those root cause(s). Why weren’t the companies asked to reveal these data.
This whole inquiry should be focused on developing a way for companies to cooperate in making these systems safer, rather than compete. The purpose of testing on public roads is to identify and uncover the challenging situations, so called “corner cases”. That means the technical description of the scenario leading up to the disengagement. Focusing on the role of the human driver(s) in the vehicle, is like focusing on the band-aide rather than the scenario that led up to you being cut. Elaine Herzberg’s death can’t be blamed on the Safety driver. She looked in the direction of Elaine 5 seconds before the crash and didn’t see anything. Uber revealed that its system detected Elaine one second earlier (6 seconds before the crash), did nothing and, obviously, didn’t report or highlight on its monitor anything like: “object/vehicle/pedestrian ahead on a collision course, do something!”. This certainly suggests that the AV system didn’t perform well.
All of the data captured by Uber during the 10 seconds leading up to this crash MUST be released to everyone so that no one’s AV system repeats the mistakes made by Uber’s system. Similarly, the 10 seconds of data captured before each and every non-routine disengagement should be published so that everyone can make their system safer. This deserves industry-wide full cooperation.
10. What data… Read more This question begins to get at the sharing of data but it asks for “safety performance” whatever that is, rather than the data from its sensors and decisions made by its algorithms in the seconds leading up to a crash of each and every non-routine disengagement. The whole purpose of testing on public roads is to understand the performance of the sensors, the data that they reported and what decisions the algorithms made in controlling the AV prior to the crash or disengagement. Having these data allows the scenario to be reconstructed and replayed in simulations and/or on closed courses to enable the root cause(s) to be identified and fixed, if possible. Because it was dark and Elaine was wearing dark clothing and her bike didn’t have side reflectors, the safety driver may not have been able to see her in time, even if she had been looking in the right direction all the time; however, the AV system did identify her as an object for 6 seconds and did nothing. Why??? What data triggered the system to classify her as an object, then a vehicle, then a pedestrian. What data led to assigning trajectories to those objects? What were those trajectories and what was done with those trajectories? A question to be asked of all of the others: here is the data from Uber, what does your driving simulator do with your algorithm? Does it crash into Elaine? What is displayed on your system monitoring screen. Would that information do you display/speak that would have caused your driver to disengage the system had the driver been paying attention to your monitor?? Alain
T. Lee, May 29, “When people think about self-driving cars, they naturally think about, well, cars. They imagine a future where they buy a new car that has a “self drive” button that takes them wherever they want to go.
That will happen eventually. But the impact of self-driving technology is likely to be much broader than that. Our roads are full of trucks, taxis, buses, shuttles, delivery vans, and more—all of these vehicles will have self-driving equivalents within a decade or two.
The advent of self-driving technology will transform the design possibilities for all sorts of vehicles, giving rise to new vehicle categories that don’t exist now and others that straddle the line between existing categories. It will also change the economics of transportation and delivery services, making on-demand delivery a much faster, cheaper, and more convenient option.
Recently we had the chance to talk to two self-driving vehicle startups that are at the forefront of these trends.
Earlier this month, the startup Drive.ai announced an autonomous shuttle service that will launch in July in the Dallas metropolitan area. The company’s vehicles straddle the line between buses and taxis—like a bus, they’re designed for shared service in a fixed area, but rather than being on a fixed route and schedule, they can be hailed on demand.
Meanwhile, Nuro is building self-driving cars for moving goods instead of people, and it recently applied for permission to test its fully driverless vehicles in Arizona. Because Nuro’s cars don’t need room for passengers—or all the safety equipment a human rider needs—Nuro’s cars can be much smaller and lighter (and therefore cheaper and safer) than a conventional car….” Read more Hmmmm…. Tim is writing about Driverless, not Self-driving, but read on. There is a lot to read. Also, his table should have Waymo, Cruise and Uber in Shared Service. While their market depth doesn’t afford them many shared ride opportunities, the only way they can become mainstream and really relevant, rather than just another polluting congesting luxury for the rich, is through ride-sharing. Alain
The End of Driving, 1st Edition: Transportation Systems and Public Policy Planning for Autonomous Vehicles
B grush & J. Niles, ISNB 9780128154519 “While many transportation and city planners, researchers, students, practitioners, and political leaders are familiar with the technical nature and promise of vehicle automation, consensus is not yet often seen on the impact that will result, or the policies and actions that those responsible for transportation systems should take.
The End of Driving: Transportation Systems and Public Policy Planning for Autonomous Vehicles explores both the potential of vehicle automation technology and the barriers it faces when considering coherent urban deployment. The book evaluates the case for deliberate development of automated public transportation and mobility-as-a-service as paths towards sustainable mobility, describing critical approaches to the planning and management of vehicle automation technology. It serves as a reference for understanding the full life cycle of the multi-year transportation systems planning processes, including novel regulation, planning, and acquisition tools for regional transportation… Read more Hmmmm…. Congratulations Bern and John. I’ve pre-ordered my copy. Discount code is ATR30. Alain
AS, May 30, “There has been a long history between AutonomouStuff and UIUC from workshops to faculty to providing products for research products. After the university purchased the AStuff Autonomy Starter Kit, we extended an invitation to graduate-level students from University of Illinois Department of Industrial & Enterprise Systems Engineering (ISE) to the Morton facilities to gather data for their graduate research projects. Each research project, unavailable to the public, questions how the world currently works and presses toward better understanding a future of autonomy…” Read more Hmmmm…. Congratulations Bobby. Alain
M. Schlinkmann, May 20, “…They’re among the 175 or so people who use ITNGateway, a nonprofit ride service for older residents in St. Charles County which expanded this month to also cover St. Louis County’s central corridor. ITN, which stands for Independent Transportation Network, is unusual because most rides are provided by volunteer drivers dispatched in their own vehicles.
In return, the drivers can get credits they’ll be able to use to pay for rides themselves when they no longer can get on the road. Another option is mileage reimbursements, although most prefer the credits.
“Those ride credits can be banked for future use,” explained Susan Kallash-Bailey, ITNGateway’s president and executive director. “Or they can donate them to a family member.” Under the program, riders can get to anywhere in their service area 24 hours a day, seven days a week for a fee of $2.50 plus $1.50 a mile…” Read more Hmmmm…. Very interesting market focus. Alain
C. Woodyard, May 29, “The latest crash involving a Tesla in Autopilot mode didn’t turn tragic, as some past ones have, but certainly was embarrassing. A Tesla Model S veered into a parked police cruiser Tuesday, severely damaging both vehicles in Laguna Beach, Calif., … The driver, a 65-year-old from Laguna Niguel, Calif., told officers that he had engaged the car’s partial self-driving system, called Autopilot. “He told us in his own statement he was in driver-assisted mode,” police Sgt. Jim Cota said.
The driver suffered minor injuries, Cota said. The parked cruiser was unoccupied, the officer standing about 100 feet away off Laguna Canyon Road as he responded to a call. Cota said the luxury electric car crashed in almost the same place as another Tesla about a year ago. The driver, he said, also pointed to the Autopilot system as being engaged. … Read more Hmmmm…. Again, what was the AEB system doing? Turned off again because it is a stationary object??? So bad!!! Alain
NTC, May 2018, “This policy paper sets out recommendations for legislative reform to: provide clarity about the situations when an automated driving system (ADS), rather than a human driver, may drive a vehicle; ensure there is a legal entity that can be held responsible for the operation of the automated driving system; establish any new legal obligations that may be required for users of automated vehicles; and outline further work that needs to be done to transform agreed policy into legislation…
The National Transport Commission is working with state, territory and the Commonwealth governments on a program of regulatory reform to ensure the Australian community can gain the potential benefits of automated vehicles, including safety, productivity, environmental and mobility outcomes. Our aim is to develop a flexible and responsive regulatory environment for the commercial deployment of automated vehicles that supports innovation and safety.
This policy paper delivers a key aspect of this reform agenda. Ministers have agreed to a ground-breaking approach to driving laws in Australia. This will see the development of purpose-built national law to allow an automated driving system to drive in place of a human.” Read more Hmmmm…. Excellent contribution from Australia who has been working diligently on this topic for some time. Alain
J. Elmerraji, May 30, “But it’s becoming increasingly clear that autonomous vehicles are the future of personal transport – and that the future is coming fast. That, in turn, is driving attention to the big companies that have equally big exposure to the self-driving vehicle trend.
One of the best ways to play the autonomous vehicle trend right now is with the chipmakers. The companies that make the “brains” behind self-driving cars own some of the most defensible intellectual property in the space right now. And beyond exposure to existing self-driving car programs, partnering with chipmakers provides a way for the scores of carmakers playing catch-up on their self-driving car tech to accelerate their pace….” Read more Hmmmm…. I consistently Buy-High Sell-Low, so you don’t want to read my views, except nVIDIA has participated in both Princeton SDC Summits. Alain
Calendar of Upcoming Events:
3rd Annual Princeton SmartDrivingCar Summit
evening May 14 through May 16, 2019
Save the Date; Reserve your Sponsorship
Photos from 2nd Annual Princeton SmartDrivingCar Summit
Program & Links to slides from 2nd Annual Princeton SmartDrivingCar Summit