In simulation on a unit G1, the policies learned under terrestrial gravity still stand on the Moon, while those trained for the Moon collapse on Earth.
On the Moon, gravity, relief, and the nature of the ground all change at the same time. Work on learned locomotion has so far dealt with reduced gravity on rigid ground or sand with Earth gravity, rarely both for a humanoid. Researchers trained twelve walking policies for a simulated Unitree G1 with the same recipe, varying gravity, terrain geometry, and contact pattern (rigid ground or granular soil), including refined Earth-to-Moon policies and others subject to random gravity.
A clear asymmetry
Earth-trained policies maintain a survival rate of 0.64 to 0.75 on flat, rigid lunar soil, while those trained or refined for the Moon are unable to stand up to Earth's gravity. The general-purpose controller SONIC, tested as-is, does not exceed 0.10 survival on the Moon. Training on rough terrain jumps the combined survival from 0.53 to 0.98 on Earth and 0.80 to 0.99 on the Moon. A randomly gravity policy trained on rough granular ground achieves at least 0.84 on Earth and 0.96 on the Moon on all key configurations, up to slopes of 20 degrees.
The authors point out that their lunar results remain simulation predictions, which depend on the contact model and the ground stiffness chosen.
Read the full article on the original source
Go to the source →