Start from the trot: behaviour cloning and residual policies

Training from scratch found small, fast steps that hobby servos may not follow. Two ways to start from the gait that already walks instead: copy it into the policy network and let PPO improve it, or keep the trot running and learn only a small correction on top. Both trained, measured on random robots and compared.

55 min

Cette page n'est pas encore traduite, elle s'affiche donc en anglais.