Cooperative Long Rope Skipping via Hierarchical Multi-Agent Reinforcement Learning

Cooperative long rope skipping requires two humanoid rope turners to manipulate a shared deformable rope while maintaining balance and synchronizing its rotation with a jumper’s rhythm. Existing humanoid sport approaches centered on individual athletic skills or independent motion replay offer limited mechanisms for handling the physical coupling through the rope and adapting its motion to the jumper’s phase. To address this challenge, we propose Marope, a hierarchical multi-agent reinforcement learning framework that combines decentralized rope manipulation with centralized rhythm scheduling. Low-level policies trained with an adversarial motion prior provide closed-loop rope control through a compact command interface, which a high-level scheduler uses to synchronize rope rotation with jumping. An integral probability metric-based diversity objective further discovers varied jumping behaviors to augment cooperative training. In simulation, Marope achieves a mean successful-jump rate of 78.7% on held-out jumping policies. We also deploy Marope in four representative real-world coordination scenarios, all of which achieve a success rate of more than 70%, showcasing the heterogeneous coordination capability of Marope across different embodiments.