Moon-walking simulations expose a one-way problem for humanoid robots

A humanoid controller trained for Earth can sometimes cope with the Moon’s weaker gravity. Training in the opposite direction proved much less forgiving: lunar-trained controllers failed even to stand when returned to Earth in a new Unitree G1 simulation study. The asymmetry matters because changing worlds changes weight, foot contact and gait together, while the robot’s inertia stays the same.[1]

Guillem Casadesus Vila, Sri Aditya Deevi and Grace Gao trained twelve simulated locomotion policies across gravity, terrain and contact conditions. The ground included rigid surfaces and deformable soil represented using three-dimensional resistive force theory. They also tested SONIC, a released generalist humanoid controller. The preprint was posted on 21 October.[1]

The transfer that works in one direction

On flat rigid ground under lunar gravity, Earth-trained policies completed between 64 and 75 percent of the tested episodes according to the study's survival criterion. The reverse was worse: policies trained for the Moon and policies fine-tuned from Earth to lunar conditions could not even complete the standing phase in Earth's gravity under the tested setup. The released SONIC generalist controller survived at most 10 percent of lunar-gravity tests. These are survival fractions in defined simulated episodes, not percentages of real missions guaranteed to succeed.[1]

The observed gaits help explain the result. In lunar gravity, controllers could rely on longer airborne phases and slower transfers of weight. Earth’s stronger downward pull makes those strategies harder to sustain. This is an interpretation of the tested controllers, rather than a universal rule for every robot.[2]

Make training rougher, not just lighter

A stronger result comes from training environments that vary terrain difficulty and gravity. Under rough-terrain training, combined-tile survival rose from 0.53 to 0.98 in an Earth condition and from 0.80 to 0.99 in a Moon condition, according to the paper's comparisons. A policy randomised across both gravities and trained on rough deformable terrain achieved seed-mean survival of at least 0.84 on Earth and 0.96 on the Moon across key test conditions up to 20 degrees of slope. Varying environments can therefore produce a controller that handles several families of disturbance rather than memorising one gravitational regime.[1]

Soil is a separate uncertainty. Lunar regolith is not simply beach sand in lower gravity: grain shape, compaction and foot pressure all affect contact. The research compared multiple contact and soil models, including held-out configurations. In sampled runs, Moon-trained policies kept much of their rigid-ground progress on simulated lunar regolith, but those results depended on the chosen stiffness and contact parameters. No actual lunar soil test was reported, and digital ground models are not a guarantee of field performance.[1][2]

The engineering insight is wider than the Moon. Robot-learning demonstrations often score performance inside the environment that generated the training experience. Real deployment changes several conditions at once: friction, load, terrain, sensor noise and unexpected contact. A controller that looks graceful in one clip may be brittle when any of those ingredients change. Testing transfer and the failure modes is more valuable than simply raising a demonstration score.[1]

The useful design target is a controller that travels between conditions. In this study, randomising gravity and making the training terrain rougher produced that adaptability more effectively than specialising for an idealised lunar walk.[1]

Report an error or suggest a correction