It was very exciting when autonomous vehicles from a handful of companies first began demonstrating the ability to go a few miles in complex traffic and pull off a few fancy maneuvers along the way. I got really excited when I saw that, and for the engineers building these systems, it did go to their heads, and timelines to scaled deployment went from being 20 years away to being 5 years away in the span of 18 months.
When deep learning and simulation were first applied huge gains were made, and the limits of these tools were not fully understood. Now that these tools are better understood, we are back to old fashioned robotics development timelines. You gotta build features, and test them, over and over, down to a very granular level. Every time you roll out into a new geography, there is a new mountain of problems to be solved. To drive the error rate down another order of magnitude, the workload goes up exponentially relative to what was needed to achieve the previous order of magnitude. It's like inverse Moore's law.
Boston Dynamics had bipedal robots balancing dynamically and walking around in the early 1980s. 35 years later, now that we have a billion times more compute to throw at the problem, well, BD's robots are still mostly just walking around, and far from being good enough to walk around in the real world. That's the real rate of progress in robotics. The prospect of autonomous vehicles has drawn billions of dollars of investment towards solving these fundamental robotics problems, and certainly that has accelerated things, but how long does the money need to keep coming in for before it starts paying out? That's anybody's guess.
BDs controllers were/are all hand-tuned using good old convex optimization. Good luck getting fancy 'Deep-RL' to work on these without multiple man years tweaking every little knob in the cost function and the algorithm.
When deep learning and simulation were first applied huge gains were made, and the limits of these tools were not fully understood. Now that these tools are better understood, we are back to old fashioned robotics development timelines. You gotta build features, and test them, over and over, down to a very granular level. Every time you roll out into a new geography, there is a new mountain of problems to be solved. To drive the error rate down another order of magnitude, the workload goes up exponentially relative to what was needed to achieve the previous order of magnitude. It's like inverse Moore's law.
Boston Dynamics had bipedal robots balancing dynamically and walking around in the early 1980s. 35 years later, now that we have a billion times more compute to throw at the problem, well, BD's robots are still mostly just walking around, and far from being good enough to walk around in the real world. That's the real rate of progress in robotics. The prospect of autonomous vehicles has drawn billions of dollars of investment towards solving these fundamental robotics problems, and certainly that has accelerated things, but how long does the money need to keep coming in for before it starts paying out? That's anybody's guess.