Event

Simulation study: transfer performance of Earth-trained humanoid controllers under lunar gravity

1 reports · 1 sources

Guillem Casadesus Vila et al. trained 12 policies in simulation on a Unitree G1 using the same training recipe, varying gravity, terrain geometry, and ground contact models (rigid or three-dimensional resistive force theory, 3D RFT), and tested Earth-to-Moon fine-tuning and gravity randomization. The study reports that gravity transfer is asymmetric. The study also tested the released general-purpose humanoid controller SONIC (with frozen parameters), but only under two gravity conditions on rigid ground. The lunar-related results are simulation predictions and depend on the contact model and soil stiffness. In sampled rollouts with a held-out soil model, the study reports that the lunar-trained policies, under low foot pressure on the G1, retained most of the rigid-ground progress on lunar regolith; excluding one seed with a shorter rigid reference, the progress ratio across seeds ranged from 0.84 to 1.39.