arXiv preprint ·
Do Earth-Trained Humanoid Controllers Walk on the Moon? Gravity, Terrain, and Soil Transfer
- Published
- Source
- arXiv preprint, posted before peer review
- Original
- Read it at arXiv
- Topics
Abstract
The Moon changes gravity, terrain, and soil at once. Learned locomotion has studied reduced gravity on rigid ground and granular terrain mostly at Earth gravity, but not humanoids under both, generalist-controller transfer, or the Earth cost of lunar adaptation. On a simulated Unitree G1, we train 12 policies with one recipe, varying gravity, terrain geometry, and ground contact (rigid or three-dimensional resistive force theory, 3D RFT), and including Earth-to-Moon fine-tuning and gravity randomization. We test them at both gravities on rigid ground and 3D RFT, and at their training gravity on three held-out soil models. We also test the frozen SONIC, a released generalist humanoid controller, at both gravities on rigid ground only. Gravity transfer is asymmetric. Earth-trained policies keep 0.64 to 0.75 survival at the Moon on flat rigid ground, Moon-trained and fine-tuned policies cannot stand at Earth, and SONIC survives at most 0.10 at the Moon. Rough training lifts combined-tile survival from 0.53 to 0.98 at Earth and 0.80 to 0.99 at the Moon. In 3D RFT, soil training helps on Earth sand, on tiles and for rough-trained policies also on flat sand. On the stiffer lunar soil it helps only flat-trained Earth policies at Earth and changes the survival of Moon-trained policies by at most a few hundredths. In the sampled rollouts of the held-out soil models, Moon-trained policies keep most of their rigid-ground progress on lunar regolith at the low G1 foot pressure, with per-seed progress ratios of 0.84 to 1.39 apart from one seed with a short rigid reference. A gravity-randomized policy trained on rough 3D RFT terrain has a seed-mean survival of at least 0.84 at Earth and 0.96 at the Moon on every key test cell up to 20 degrees. Lunar results are simulation predictions conditional on the contact model and soil stiffness.
The authors’ abstract, as arXiv lists it. arXiv’s listings, abstracts included, are free for anyone to reuse.
Read it on ReadSearch
Each opens after you create an account, and brings you back to this paper.
More in Robotics
PHD-Hand: A Printable Humanoid Dexterous Hand with Linkage-Driven 3-DOA Digits
Know When to Escalate: Critic-Guided Adaptive Execution for Cloud-Local Embodied Agents
Causal world models under physical parameter shifts
OASIS: Online Anchor-centric Scene Inference System for Open-Vocabulary Retrieval and Navigation in Dynamic Environments
Learning to Absorb Corrections: Critic-Weighted Residual Flow for Vision-Based Insertion
Get the few papers worth your time in Robotics every morning.
ReadSearch reads the new papers in your field each day and brings you the few worth reading, each with its key finding in a sentence.