Learning Goal-Following Locomotion Controllers for Humanoids Using Demonstration and Reinforcement Learning

Kishor Kumar, Kameshwar Rao, Somdeb Saha, Vighnesh Vatsal, Kaushik Das

Keywords: Humanoid robots, Reinforcement Learning, Imitation Learning

Abstract: Humanoid robots must coordinate locomotion with upper-body motion while responding to high-level goals. This work presents a goal-conditioned controller trained through Archive of motion capture as surface shapes (AMASS)-based imitation learning and reinforcement learning (RL) in the LocoMuJoCo framework. A policy is first pretrained on AMASS trajectories to acquire human-like gait dynamics and coordinated arm–leg motion, then fine-tuned with RL to track target root velocities and hand poses using a lightweight DeepMimic-inspired reward. Using a Unitree H1–scale model, we find that AMASS-initialized RL converges faster and yields higher stability, smoother motion, and more accurate goal tracking than RLfrom-scratch. These results demonstrate an effective and scalable strategy for developing natural whole-body humanoid control suitable for future loco-manipulation tasks.

Submission Number: 16