Image credit: Chenjia BaiAbstract
While recent advances have demonstrated strong performance in individual humanoid skills such as upright locomotion, fall recovery and whole-body coordination, learning a single policy that masters all these skills remains challenging due to the diverse dynamics and conflicting control objectives involved. To address this, we introduce X-Loco, a framework for training a vision-based generalist humanoid locomotion controller. The key insight is that the fundamental locomotion skills of standing up, recovering from falls, walking, running, and jumping share a common latent representation. We propose a novel policy distillation approach that synergistically transfers knowledge from independently trained expert policies into a single multi-skills policy. To tackle the conflicting objectives arising from distinct skill dynamics, we introduce a skill-conditioned adaptive loss weighting mechanism that dynamically balances the optimization process. Additionally, we integrate a domain randomization strategy specifically tailored for the distillation process, enabling sim-to-real transfer without requiring any real-world fine-tuning. Extensive experiments demonstrate that our unified policy successfully masters diverse locomotion skills in both simulation and the real world.