SwingBot: Learning Whole-Body Brachiation for Humanoid Robots

SwingBot presents a reinforcement-learning framework for continuous brachiation on a high-DoF humanoid using passive wrist hooks. The controller operates through a 20-DoF policy interface and must coordinate release, whole-body swing, alternating contact, and capture under partial observability. Its main innovation is residual keyframe guidance: sparse biomimetic postures make release–swing–capture transitions reachable during early PPO exploration, then the guidance is annealed away over 2000 epochs. A recurrent state-space model, or RSSM, estimates hidden segment-relative displacement and left/right hook contact from proprioceptive and action histories; its 64-dimensional deterministic latent is available during deployment, while privileged simulator variables train the critic and world model. In simulation, adding the RSSM increased all-8 traversal success from 57.28% to 61.27% with 1.5-second command switches, and from 43.12% to 48.86% with 1.0-second switches; removing keyframe guidance caused complete failure across the tested seeds. Hardware training used Isaac Lab with NVIDIA Isaac Sim, and the robot demonstrated alternating bar traversal, disturbance recovery, payload carrying, and bar-spacing variation. Under nominal real-world conditions, single-swing success reached 92.86% and all-8 completion reached 60.00%. The system also carried a 1 kg payload, although repeated transitions were limited by shoulder-motor heating, and future versions will need visual perception and active grippers for arbitrary bar layouts.