Researchers generate humanoid control networks that perform real-world motions

Researchers report in a preprint that a generative model produced control networks capable of making a Unitree G1 humanoid walk, dance and perform a spin kick in the real world. They trained the system on networks collected through reinforcement learning for that robot. An autoencoder compresses the control networks' numerical parameters into a smaller representation, and a diffusion model learns to generate those representations. The researchers also found that first-layer parameters reflected the robot's physical structure under both reinforcement and imitation learning. The work demonstrates feasibility for generating humanoid controllers, with the reported deployment results tied to the Unitree G1.