Rich's Robotics Blog
May 2, 2022 Check out my new guitar
I can't believe it's been more than two years since I last posted
here. When I look at my last posts about self-driving cars I can
definitely say that nothing has happened in the last two years that
makes me think self-driving cars are going to be ubiquitous before
When covid started a little more than two years ago, I decided to
learn to play guitar with the isolation time at home and I spent an
hour or two every day practicing. Soon, my back hurt from the weird
position typical electric guitars put your back in when you set them
on your knee to play sitting down, Of course, I decided I needed to
build a better electric guitar and I think it turned out great. Here
is the story of my ergonomic guitar
Feb 22, 2020 Fully Automated Self-Driving Cars Will Be
Available in 2068!!!
I’ve previously written today’s state of the art in self-driving cars reminds me very much of a 16-year-old with a newly-minted driver’s license. This is nothing provable in a strict sense, but I find the analogy interesting. This follow-up article carries the analogy with humans four decades into the future of self-driving cars. Of course, nothing can prove the future, but the analogy can help us prioritize the development of self-driving cars in a way that is most beneficial to us people, and as we have heard many times, saving humans from the horrors of car crashes is the primary reason for self-driving cars.
This starts with a question about when self-driving cars will achieve a vision. Before answering this question, we need to know what that vision looks like. I will make that simple by using my own personal vision for self-driving cars. Here it is. Self-driving cars should be able to drive to my house, pick me up and then safely drive me to where I want to go. By safe, I mean at least as safe, and hopefully much safer, than if I drove myself. In my vision, self-driving cars are ubiquitous and share the roads with human-driven cars.
Now we have the current state of the art in self-driving cars and a vision for the future state. What’s left is projecting how long it will take to get from state A to state B. For that, we need to do some more work. If 16 isn’t old enough in equivalent human years for self-driving cars to achieve the vision, then how old is old enough? It is well known that 16-to-19-year-olds have the highest rates of accidents where people are injured or killed. We should stay out of that age range. Therefore, I say that self-driving cars need to get to 19 human-equivalent years of maturity. That leaves 3 more human-equivalent years to go.
This leaves the question of mapping human-equivalent years into self-driving car development years. We know it is going to be much more than 3 years before self-driving cars are unleashed on the general roadways, so it is more than 1-to-1, but does 1 human-equivalent year equal 3 self-driving car development years? 5? 10? 20? Fortunately, we have a data point in the past to help answer that question. That data point is the 2004 Darpa Grand Challenge. You can find the rules to that year’s challenge online, but in summary, the cars had to drive over desert roads while following a rudimentary electronic map without human assistance and without traffic on the roads.
At the time of the 2004 challenge I wrote, “…the task itself seems so simple. Any human with a $5,000 ATV could do it.” Looking back, I would say these cars were about as good as a 15-year-old human that just got their learner’s permit. It’s been 16 years since that Darpa Grand Challenge and self-driving cars have advanced about 1 human year. That means a human-equivalent year in self-driving car development time equals 16 calendar years. Multiplying this by the three human equivalent years to go gives 48 calendar years before self-driving cars will achieve my vision.
At this point I know you are thinking 42 years is a very long time. There is no way it is going to be 42 years before self-driving cars are ubiquitous on our roads. To that I say, not so fast. The self-driving car folks have already been at it 14 years. That’s an eternity in technical revolution time. Google went from a dorm-room to world-domination in less than 10 years (Google was founded in 1998 by Larry Page and Sergey Brin while they were Ph.D. students at Stanford University in California. An initial public offering (IPO) took place on August 19, 2004).
Furthermore, the self-driving car companies have been riding the wave of ever-increasing computing power over the last 14 years. Unfortunately, the semiconductor revolution is getting long in the tooth. Moore said years ago that Moore’s law no longer holds true. Individual microprocessors are not getting much faster any more. Of course, computers are increasingly networked and that increases aggregate computing power by summing over many individual computers. Unfortunately, self-driving cars do not have the luxury of “going to the cloud” when they must make critical decisions. These decisions need to be made in tiny fractions of seconds. Going to the cloud for computing power is not going to do the trick for real time decision making when the vehicle is going sixty miles per hour.
Finally, an unfortunate reality putting the brakes on the adoption of self-driving cars, is that changes in the automobile industry move at a snail’s pace. The original Prius debuted on the market in 2000. That’s 18 years ago. Today just 2% of car sales in the US are hybrids and electric-only vehicles are a smaller percentage than that. I know the potential to save lives should be pushing the self-driving car technology, but the potential to save the world isn’t pushing the adoption of hybrid or electric cars all that fast.
Feb 12, 2020 What is the state of the art in self-driving cars?
To answer this question, at least at a macro level, I look at two examples of self-driving car projects, one in industry and one in academia. In the self-driving car industry, there is nothing more state of the art than Waymo. We’ll start there. After that we’ll look at a recent project conducted by the Computer Science and Artificial Intelligence Lab, Massachusetts Institute of Technology (MIT) and presented at the 2018 International Conference on Robotics and Automation. It doesn’t get much brighter than that in the research community.
As is widely known, Waymo originated as a project at Google in 2009 and eventually became its own subsidiary. Waymo has also made no secret of its plan to roll out an autonomous ride-hailing business in Phoenix, Arizona in late 2018. In preparation for the roll out, Waymo is amassing enormous databases filled with real-world experiences recorded during actual driving experiences on the Phoenix streets. In addition to these recordings, the company has also created detailed maps the cars reference while they drive. I’m sure anyone that has lived in the area can attest that Phoenix has almost perpetual clear driving weather. It only gets about 8 inches of rain per year and the last time they had measurable snow at the Phoenix Sky Harbor International Airport was in the 1930’s.
I am sure Waymo is confident their cars will operate safely, but don’t forget they have greatly constrained the environment these cars will operate under. The cars will be operating in good weather and on roads with which they have a lot of experience driving. Does that sound familiar? It does to me. Those were the rules when my kids got their driver’s licenses at 16. They could only drive on very familiar roads and then only in clear weather. Why did my wife and I constrain their driving like this? Because that was what we thought was safe. At least from this perspective, Waymo is about at the driving prowess of a 16-year-old with a freshly-minted license.
The next data-point comes from the research community, specifically MIT. This group wrote about their project in a paper titled, “Autonomous Vehicle Navigation in Rural Environments without Detailed Prior Maps.” As I mentioned above, the Waymo cars continuously reference detailed maps while they drive, and detailed means detailed. These detailed maps know where the stop signs are, the traffic signals, construction zones, etc. They are far more detailed than the kinds of maps most of us are familiar with. The MIT group, on the other hand, did not use these detailed maps because they do not exist for most roads. Their goal was to develop technology enabling self-driving cars to operate with access only to widely-available, “standard-definition” electronic maps. To that end, the MIT group used maps from the Open Map Project, which are very much like Google or Apple maps.
And that’s what the MIT group demonstrated. Their research and development demonstrated that self-driving cars could drive on rural roads while following standard electronic maps. I will also note that this demonstration was in a highly-constrained environment. The demonstration operated the cars in clear weather on rural roads without much, if any, other traffic around. The cars could operate safely on open, rural roads while following electronic maps. I would say that, too, is about at the skill level of a 16-year-old when they first get their driver’s license.
I’m guessing the self-driving car companies would feel this analysis under-estimates the prowess of their self-driving cars. They would tell you that the real problem is the 16-year-old human drivers on the road. That may be so, but it will be many decades (>5) before human drivers are off the roads. Besides, I haven’t really presented an analysis. This is more of an observation. It is based on two state of the art data points from the self-driving car world, my own experiences teaching my kids to drive, and framed by the perspective of 30+ years in the field of robotics and automation. The state of the art in self-driving cars reminds me of my kids driving when the first got their licenses at 16.
July 11, 2019 Overselling technology
I just rewatched an old TED talk on YouTube where a guys talks about Google’s driverless cars. The video is interesting enough.
The guy does a good job of talking about how hard the problem is. Unfortunately, at the end he makes a completely self-serving proclamation that his team’s goal is for his son not to need a driver’s license in 4.5 years. Guess when the TED talk was published? June 26, 2015. Does anyone think that people are not going to need driver’s licenses within the next half-year? Of course not. One of the things I’ve learned over the years is the more salesmen oversell their technology, the further they are from solving the problem at hand. Driverless cars are nowhere near ready to make cars with drivers obsolete.
I'm going to write more about this.
October 11, 2018 Baxter bites the dust!
was one of the two original "Cobot" robot manipulators. A
company called Rethink Robotics out of MIT designed, manufactured
and sold Baxters. Universal Robots is the other original Cobot
company. Cobot abbreviates collaborative robots. I will agree these
were the first cobots, since that's when the word was invented, but
I certainly don't agree these were the first examples of companies
making collaborative robots. I'm going to write a new page about
Cobots and collaborative robots, but for now here is a brief thought
on Baxter that isn't getting as much press.
Baxter was a piece of junk! It was so bad mechanically
that it was incapable of all but the most trivial tasks. The
promotional videos for Baxter would show it moving plastic cups from
a conveyor belt to a box. Not because that is impressive, but
because a plastic cup was all Baxter could lift! I felt sorry for
the thing. It shook so badly when it moved that I could barely
watch. People at trade shows laughed at it. When these crummy
mechanicals were being designed, the computer scientists at Rethink
thought their control algorithms would take care of the issues. This
is an arrogant misconception among many engineers in the computing
disciplines. Once crummy mechanicals have been built into a robot,
that robot is doomed. I don't care how fancy the control algorithms
Baxter was also promoted as being inherently "safe,"
though that is NOT generally true of Cobots. This was true of Baxter
only because it was too weak and too slow to hurt anything, which,
of course, also made it unable to DO anything (except move cups
around slowly). The notion that a robot has to be weak and slow to
be safe is pathetic. I'm glad Baxter is gone. Good riddance to bad
September 25, 2018 Are the self-driving car companies that
concerned about saving lives?
Watch Chris Urmson’s TED talk or listen to just about anything from the self-driving car companies and you will hear heart-felt
attestation of their concern for human safety. In his talk, Chris says, “if we can just cut the accidents in half” with self-driving cars it will be all worth it. Of course, that means the public must wait for the roll-out of driverless cars to get those safety benefits. This is a patently false precondition. As I will discuss below, we can cut traffic fatalities in half with existing technology. Furthermore, it will be far less expensive and can be deployed much more quickly than driverless cars.
Why aren't the self-driving car companies pushing solutions like
these if they are so concerned about saving humans from the horrors
of car crashes?
Let’s look at the top three causes of traffic fatalities, based on data from the Insurance Institute for Highway Safety and the National Highway Transportation Safety Administration.
1. Speeding – 30.6%
2. Drunk Driving – 31.1%
3. Distracted Driving – 16%
I’ll now discuss these in order.
Speeding – Eliminating virtually all accidents associated with speeding is entirely doable with existing technology. Simply enforce speeding laws automatically. Most driving in this country, and I’m going to guess at least 90% now happens in areas that are electronically mapped. The speed limits where cars are driving are known. Cars should be programmed to ignore requests by the driver to exceed the posted speed. We could allow drivers to exceed the posted speed briefly for safe passing and the like, but that’s it. AS A FIRST STEP, all cars sold in this country should be equipped with a governor that will not allow the car to go faster the 85 mph. As I’ve written before, why on earth will my wife’s Honda go 120 mph? It is not legal to drive that fast in this country anywhere. Why is it legal to sell a car that will go that fast? Machine guns are illegal everywhere in this country and it is illegal to sell machines anywhere in this country. The same should be true for cars capable of exceeding 85 mph.
Automatic control of speeding is something the self-driving car companies should be able to get behind. One of the visions behind the self-driving car future is that cars will be networked. They will know each other’s driving paths and intentions and use this information to eliminate accidents and efficiently utilize driving surfaces. Networking cars like this requires automatically communicating real-time data to cars, and the cars then responding to the data. How about as a first-step, we start communicating speed limit data to cars and the cars can respond by ignoring driver inputs that ask it to exceed the speed limit?
Drunk driving – I don’t know that all drunk and drugged driving can be eliminated as long as people are driving cars, but existing technology can certainly substantially reduce the problem. Drunk drivers (and drivers on other drugs) exhibit known behaviors while they drive. For example, they swerve in their lanes. Police are trained to spot these behaviors. The “brains” in existing cars can also be taught to identify these markers. It’s not clear to me at this point what the car should do after it has identified a potentially impaired driver. Maybe limit the speed to 25 mph? I would also be fine in the car notified the police.
Distracted Driving – The easiest one to attack here is smart phones. User input to smart phones should be disabled in moving cars. This is 100% doable by downloading new software into the phones. We don’t even need to change hardware. We should demand the smart phone companies do this immediately. Unfortunately, people seem uninterested in doing that, but eventually people should get tired of their loved ones being killed because of
smart phones. I suppose that could generate a groundswell of support that would force the issue. More likely, it will be resolved in
litigation and then government decree.
There we go. If we do these three things we can cut driving fatalities in half. Why aren’t the self-driving car companies screaming about this? Is it because they care more about making dollars than saving
lives? I'd like to see them prove that wrong by supporting simple
changes that will save tens of thousands of lives in the
interim while they are developing their self-driving car
September 13, 2018 Darpa Robot Race Begins "At stake is a $1 Million first prize!"
I can't help re-running the paragraph I wrote below about the
2004 Darpa grand challenge. I was much more pessimistic than most
people writing about robotics technology in those days, as I am now.
Guess what? I was right about no team finishing the 2004 challenge
and I will proven right that there will never be a robot with a
brain built on a semiconductor-based computer. Of course, by
"robot" I'm writing about something that might be mistaken
for a human (as Capek intended the word to be understood), not the
myriad of devices and algorithms people call "robots" to
make them sound cool. Seriously, people will call anything a robot
to get attention.
(From March 11, 2004) The publicity for the this race has been brewing for months. The basic challenge is this: a fully-autonomous (no human control) car has to travel a fixed-course 250 miles through the Nevada desert. The quickest time that also averages at least 15 mph wins. As of now, the favorite in the race is Red Zone Robotics (associated with Carnegie Mellon University in Pittsburgh), though my bet is that none of the cars will make it. The Red Zone team has spent roughly $3 million developing their entry; so they are clearly not in it for the prize money. What I find interesting is that the task itself seems so simple. Any human with a $5,000 ATV could do it. This clearly illustrates just how far robotics has to go in terms of approaching human intelligence. I'm sure the Red Zone entry has more sensor and computing power than anyone will be able to cram into a human-scale robot within the next ten years. Marshall Brain predicts that by 2050 we will have $10,000 robots with computing power roughly on-par with humans. He may be right about the raw computing power, but it's contests like this Darpa challenge that make me believe that those robots will have no where near a human's ability to cope with unstructured environments.
September 11, 2018 Why do I care about self-driving cars?
Self-driving cars are not robots, at least according to the original meaning of the word, so why do I care about the state of the art of in self-driving cars. The answer is that robotic cars are a very “clean”
robotics application, at least in terms of the mechanicals and lower-level electronics. The task of making the physical structure of a robot with anywhere near the manipulation and locomotion capabilities of a human is extremely hard. I believe it could probably be done using existing technologies, but that is not a given, and it would take a huge
investment of time and money.
Regardless, self-driving cars don’t have this problem. “Drive by wire” control systems for cars have been common place for years. There’s no need for physical manipulation of steering wheels or brake pedals. Those signals can be sent electronically. Cars are pretty big, at least relative to humans. There’s plenty of room for sensors and computers and the like. There’s also plenty of power, either with giant (again relative to humans) batteries or internal combustion engines.
Self-driving cars really distill the problem to the functioning of the robot’s brain. That’s what makes it so interesting to me. Driving a car is one of the easiest things people do. Extending that
task to anything like human intelligence must be at least ten times harder and probably 100 times harder, or more. If we can’t even make the brains of self-driving cars, why would anyone think we could make the brains for a robot?
September 10, 2018 Why are robotic cars so hard?
On September 25, 2012; Senate Bill 1298 “Vehicles: autonomous vehicles: safety and performance,” which establishes safety and performance standards for cars operated by computers on California roads and highways, was filed with the California Secretary of State. Later that year when the bill was signed into law, Sergey Brin declared, “"You can count on one hand the number of years it will take before ordinary people can experience this.” It’s been more than five years since Mr. Brin made this declaration, and obviously he was optimistic.
Chris Urmson, a leading technologist from the early days of robotic cars, declared his son wouldn’t need a driver’s license because robotic cars would take care of the driving by the time he was 16. Dr. Urmson’s son turns 16 next year. Does anyone think robotic cars will be taking care of the driving by next year?
Here’s a recent quote from Bryan Salesky, another well-known technologist from the early days of robotic cars and the current CEO of Argo AI, “We’re still very much in the early days of making self-driving cars a reality. Those who think fully self-driving vehicles will be ubiquitous on city streets months from now or even in a few years are not well connected to the state of the art or committed to the safe deployment of the technology. For those of us who have been working on the technology for a long time, we’re going to tell you the issue is still really hard, as the systems are as complex as ever.”
What is making replacing humans in cars so hard? Modern cars are already “drive by wire,” so it’s not that we need humans to physically manipulate controls inside the car. Is it a human’s natural sensing superiority over artificial sensors? Absolutely not. Modern robotic cars have cameras, radars, lidars, altimeters, gyroscopes, accelerometers, ultrasonic rage finders, GPS and more. They have far more raw sensing capabilities than humans.
There is no disputing that engineers can build the electronics and mechanicals of robotic cars and fill them with sensors that would put a human to shame. That leaves the “brains” of the car as the missing
link. That’s the part that has me worried. I’ve written many times before that there is no path to human intelligence in a digital computer.
The development path of robotic cars is making me feel correct. The
big tech players have poured tens of billions of dollars into the
problem, yet robotic cars have become beholden to ever-increasing computing power
to enable their technical solution. Unfortunately, the digital revolution is
very long in the tooth. No serious person believes we are going to
keep seeing "Moore's Law" kinds of advances in digital