← All writingFirst posted ·
Humanoid SRL
What can a humanoid do with a third arm? We integrate a supernumerary robotic limb in two configurations, mounted at the neck or back, to extend the robot’s reach and provide additional support while its native arms are occupied.
From feasible motion to hardware
Trajectory optimization provides whole-body motion references. Learn-to-Teach reinforcement learning trains a feedback controller that tracks these references and transfers from simulation to hardware without additional training. The experiments include carrying objects through doors, stabilizing a box stack, and transporting a table with a human partner. Vision-guided diffusion policies extend the controller to selected autonomous tasks.
Publication
arXiv preprint, 2026