Publication

Training Humanoid Robots to Walk in Lunar Gravity Using Reinforcement learning

PUBLICATION DATE: 12 December, 2025 PUBLICATION AUTHOR/S: Klein, Benjamin & Roberts, Jonathan

Humanoid robots are promising candidates for lunar exploration, but adapting locomotion to reduced gravity remains a challenge. This paper investigates the use of reinforcement learn ing to train a Unitree G1 humanoid to walk in simulated lunar gravity using NVIDIA Isaac Sim 5.0 and Unitree RL Lab. Three policies were developed: an Earth baseline, a fine-tuned lunar model adapted from the Earth policy, and a lunar-from-scratch model. Results show that lunar policies autonomously discovered gait adaptations, including shortened strides, toe-dominant walking, and a forward shifted centre of mass, consistent with prior low-gravity analyses. Fine-tuning proved more efficient, achieving performance comparable to training from scratch in less than half the time.

These findings demonstrate that reinforcement learning can effectively adapt humanoid loco motion to extraterrestrial conditions and high light fine-tuning as a practical strategy for planetary robotics missions.

RELATED PROGRAM/S:
Biomimic Cobots Program Postdoctoral Researcher based at UTS Publication link

View all publications