Learning Skateboarding for Humanoid Robots through Massively Parallel Reinforcement Learning

TL;DR

The paper addresses enabling skateboarding with a full-size humanoid robot using reinforcement learning. It extends a periodic-reward locomotion framework to skateboarding on REEM-C and leverages massively parallel training in Brax/MJX to learn a forward-velocity skating policy with deck-state awareness. The contribution includes a detailed simulation setup (12 actuated leg DoF, four-wheel deck in MJX), a periodic-gait RL formulation with deck-tracking rewards, and large-scale PPO training demonstrating a pushing-skateboarding behavior with balanced dynamics. The work suggests a viable path toward transferring skateboarding skills to real hardware and expanding to glide and turning phases, highlighting the potential of periodic rewards and parallel simulation for humanoid locomotion.

Abstract

Learning-based methods have proven useful at generating complex motions for robots, including humanoids. Reinforcement learning (RL) has been used to learn locomotion policies, some of which leverage a periodic reward formulation. This work extends the periodic reward formulation of locomotion to skateboarding for the REEM-C robot. Brax/MJX is used to implement the RL problem to achieve fast training. Initial results in simulation are presented with hardware experiments in progress.

Figures (2)

• Figure 1: REEM-C on skateboard (left), skateboard (right).
• Figure 2: REEM-C skateboarding forward at 0.4 m/s