Excellent video. Dave Lee is looking for some robotic experts to interview. Maybe that can help your channel out.
Tho I would like to know in your opinion where Tesla is taking this using the FSD software stack? Why are they spending so much time doing all the classical stuff right now? And you mentioned that deep learning also doesn't translate very well into the real world, what can potentially Tesla bring to the table that can make it work?
Thanks. I'm not trying to market my channel nor make money - just get my views expressed.
I think it's great they are using the FSD software stack for perception - this is a major obvious advantage and "cheat code" to start off with that others don't have.
My guess is much like the FSD planning code was (and still partially / mostly is?) Software 1.0 that they want to convert to Software 2.0 (neural nets), the same approach is happening with the bot. They want to get some code up and running to get it "working" at a very basic level first. Plus they wanted to have a demo very soon after having the prototype built.
Yes I think they should stop spending so much time on classical stuff and move to deep learning, but it may take years to get that working well.
Deep learning can work well, it will just have to eventually be learning on real world data with appropriate fundamental cost functions, not trying to match specific human or simulation - generated trajectories. Fundamental cost functions like - 1) achieve task without falling down, 2) keep energy expenditures low, 3) avoid high impact forces. Things that humans also tend to optimize for when advancing their motor coordination.
Step 1: Deep reinforcement learning in simulation on a large variety of basic balance / movement tasks and diverse challenging conditions, with core cost functions mentioned above (
not following trajectories - maybe only in a very loose sense). They could have this done in 2023.
Step 2: Deploy as many robots as they can into a special robot test facility. Facility is designed to allow robots to move around tethered from above and interact with all different types of surfaces and perturbations. Robots are initialized with motor code developed in Step 1 but will still fail initially (like the facebook research). Robots will iteratively improve by sending failure / success data to cloud and receive updated code to further iterate while improving the same global cost functions.
[note, we could avoid the "tethering" if the bots can really learn how to fall well (and get back up) first - we simply don't want to damage the bot]