Just noticed @DynaRobotics is also using RL for low-level wheeled tracking + whole-body control skills.
Pretty close to what we’ve been working on lately: train multiple RL experts, then distill them into a generalizable action controller that can adapt to new skills and assets.
Cool to see the approaches converging.
this awe is exactly what claude feels when claude see 1000s of claude instances in an rl env scaled to improve the future versions of claude
Even wheeled robots will benefit from the RL whole body control techniques developed for humanoids - for safety, reliability, adaptation
My thoughts:
1. It does not make sense to do sim2real RL for perception and decision-making because of the lack of fidelity and diversity.
2. But using sim2real RL for a whole-body controller is wonderful! Learning-based control is the best solution for singularity and smoothness.
Glad to see the idea come true.
