Achieving Stable High-Speed Locomotion for Humanoid Robots with Deep Reinforcement Learning

Humanoid robots exhibit remarkable potential in human-centric environments, offering advantages rarely attainable by other robotic forms. However, achieving stable high-speed locomotion remains a critical challenge that further limits the applicability of humanoid robots in realistic scenarios. This difficulty arises primarily from a high CoM, difficulty in coordinating upper-body joints, and the neglect of kinodynamic constraints. To address this, a novel method, Kinodynamicconstrained Stable Locomotion Control (KSLC), is proposed, integrating deep reinforcement learning with kinodynamic priors. KSLC promotes coordinated arm movements to counteract destabilizing forces, thereby enhancing whole-body stability. In addition, velocity-related reward functions combined with curriculum learning strategies are incorporated into policy training to further improve velocity tracking performance. Comprehensive experiments demonstrated that KSLC enables humanoid robots to accurately track velocity commands up to $3.5 \mathrm{m} / \mathrm{s}$ with significantly reduced gait fluctuations compared to the baseline. Moreover, sim-to-sim cross-validation in a high-fidelity environment confirms the robustness of KSLC, underscoring its potential for real-world deployment. The experimental video can be found in the supplementary materials.