Massively Parallel Sim-to-Real for Humanoid Locomotion
In 30 seconds
RL policies trained in GPU simulation transfer zero-shot to a commodity humanoid on rough terrain.
Research question
Does domain randomisation scale to full humanoids?
Problem
Humanoid balance is brittle under sim/real gaps.
Previous approach
Model-based controllers with hand-tuned gains.
New approach
Terrain curricula plus actuator-network modelling.
Results
Stairs, slopes and pushes on a Unitree platform.
Limitations
Locomotion only; no manipulation.
Industry impact
Lowers the barrier for low-cost humanoids to walk reliably, compressing hardware differentiation.