Booster Lab: A Data-Centric Pipeline for Learning Deployable Humanoid Locomotion Policies

2026-06-26

AI Summary

This work addresses the scarcity of high-quality, physically feasible, and transferable motion data that hinders locomotion learning in humanoid robots, exacerbated by mismatches between human demonstrations and robot morphology, low fidelity of open-source datasets, and insufficient physical realism in simulated trajectories. To overcome these challenges, the authors propose a data-centric, end-to-end training and deployment framework that integrates cleaning and alignment of multi-source heterogeneous motion data, real-to-sim dynamical adaptation, adversarial motion prior (AMP)-based reinforcement learning, and an efficient sim-to-real transfer strategy. The approach enables natural and stable locomotion control on the Booster T1 platform and demonstrates strong cross-platform generalization on the K1 robot, significantly enhancing policy feasibility, naturalness, and transferability.

Abstract

Humanoid robot motion learning requires not only task-oriented control policies but also physically feasible and natural behaviors that can be transferred to real robots. However, robot-feasible motion data are often scarce: raw human demonstrations may be incompatible with the robot morphology, open-source clips vary in quality, and simulation-collected robot trajectories still require feasibility checking. To address these challenges, we propose a data-centric training and deployment pipeline that integrates motion data curation, real-to-sim model adaptation, AMP-based reinforcement learning, and sim-to-real deployment. We validate the framework on the Booster T1 robot and further provide preliminary cross-platform validation on Booster K1.

Problem

• humanoid locomotion
• motion data scarcity
• robot feasibility
• data compatibility
• sim-to-real transfer

Innovation

• data-centric pipeline
• real-to-sim adaptation
• AMP-based reinforcement learning
• sim-to-real transfer
• humanoid locomotion