Comment by cubefox
5 days ago
That would be like trying to pretrain GPT-1 from synthetically generated data only. It probably wouldn't work because the synthetic data doesn't resemble real world data enough.
It did work for AlphaGo Zero (and later AlphaZero), which were entirely trained on synthetic data. But that's for very simple games with strict formal rules, like Go and chess.
A kinematic model of the robot is a physics simulation of the robot. I don't see why that wouldn't resemble real world data enough.
Not just the robot has to be simulated, the entire part of the world it interacts with also has to be. Even the most realistic video games resemble actual videos of the real world only very superficially.
Most realistic video games don't simulate all of the physics required. Even if we just stick to simulating the motion of the robot itself in an empty space, all of that data can be generated synthetically once at the appropriate precision and reused many times, just like training data for LLMs.