Massively Parallel Sim-to-Real for Humanoid Locomotion
Noch nicht auf Deutsch verfügbar: Anzeige im englischen Original.
In 30 Sekunden
RL policies trained in GPU simulation transfer zero-shot to a commodity humanoid on rough terrain.
Forschungsfrage
Does domain randomisation scale to full humanoids?
Problem
Humanoid balance is brittle under sim/real gaps.
Bisheriger Ansatz
Model-based controllers with hand-tuned gains.
Neuer Ansatz
Terrain curricula plus actuator-network modelling.
Ergebnisse
Stairs, slopes and pushes on a Unitree platform.
Grenzen
Locomotion only; no manipulation.
Bedeutung für die Branche
Lowers the barrier for low-cost humanoids to walk reliably, compressing hardware differentiation.