Co-jump: Cooperative Jumping with Quadrupedal Robots via Multi-Agent Reinforcement Learning

Researchers developed Co-jump, a multi-agent reinforcement learning system enabling two quadrupedal robots to synchronize jumps far beyond solo capabilities. Using MAPPO and a progressive curriculum, the robots achieved precise coordination solely through proprioceptive feedback, with no explicit communication. In hardware tests, one robot reached a 1.1-meter foot-end elevation—a 144% improvement over a standalone robot's 0.45-meter jump—and the pair successfully jumped onto platforms up to 1.5 meters high.