▽Mimic and NablaEngine for Efficient Humanoid Control

Researchers Weyl Lu and Yubei Chen have introduced ▽Mimic, a first-order policy learning framework for humanoid control that uses exact gradients from a differentiable simulation engine called NablaEngine. By bounding rare unbounded samples with median-relative gradient clipping and reusing costly simulated windows through multistep regression, ▽Mimic achieves high scores in humanoid motion tracking significantly faster than standard zeroth-order methods or baseline first-order approaches. Open article →