SwingBot Makes Humanoid Brachiation Learnable With Guided Reinforcement Learning

Fudan researchers combine motion scaffolds, hidden-state estimation and passive wrist hooks to teach a humanoid alternating overhead-bar handoffs.

SwingBot is a research demonstration, not a commercial robot. In an arXiv paper, a Fudan University team reports continuous brachiation—releasing one bar, swinging, capturing the next and alternating hands—on a physical high-degree-of-freedom humanoid.

The engineering problem is not simply generating a swing. The controller must time contact changes while coordinating the legs, arms and body momentum. SwingBot first makes that sequence discoverable. During early reinforcement-learning training, biomimetic keyframes guide the robot through release, swing and capture postures. The scaffold is gradually removed, leaving the learned policy in control.

The second piece handles information the physical robot cannot directly measure reliably. The deployed policy uses onboard proprioception—joint and body-state measurements—from four recent frames, plus a recurrent state-space model (RSSM) latent. The RSSM estimates compact clues about body displacement within the current bar segment and whether either wrist hook is in contact. In simulation, a critic can use those variables directly; on hardware, the actor must infer them from history.

The controller commands 20 degrees of freedom: 12 in the legs and eight in the arms. It sends residual joint-position targets to low-level proportional-derivative controllers. The Fudan team reports complete alternating hand sequences on a physical bar, alongside tests with an 11-kilogram payload, external disturbances and different bar spacings. Its quantitative evaluation used five continuous trials per condition, with each planned for up to eight swings.

The demonstration remains bounded by the hardware and training setup. The policy has no external perception, so bar spacing must stay within the range covered by training and testing. Repeated traversal can also heat the motors and reduce available torque. Endurance testing and visual perception are the practical next steps before this becomes a field locomotion capability.