Humanoid control, generative modeling of policy weightsExpert
Modeling and Understanding Neural Weights of Humanoid Policies with Diffusion
Berkeley researchers trained a diffusion model on thousands of finished humanoid controllers and used it to generate brand-new controller networks, which then walked, danced and spin-kicked on a real Unitree G1 without fine-tuning. The preprint also finds that the first layer of every such network encodes the robot's body structure, whether trained by reinforcement or imitation learning.