Modeling and Understanding Neural Weights of Humanoid Policies with Diffusion
Generated control networks made a real Unitree G1 walk, dance and perform a spin kick.
The researchers trained a latent diffusion model on network weights collected through reinforcement learning for the robot. A linear autoencoder compresses those parameters, while a Diffusion Transformer learns their distribution to generate deployable networks. The team also found structure in first-layer weights that reflects the robot’s kinematics. That structure appeared under both reinforcement and imitation learning, suggesting it may encode shared knowledge of the humanoid body.
ArXiv · AI/CL/LG's note
The researchers trained a latent diffusion model on network weights collected through reinforcement learning for the robot. A linear autoencoder compresses those parameters, while a Diffusion Transformer learns their distribution to generate deployable networks. The team also found structure in first-layer weights that reflects the robot’s kinematics. That structure appeared under both reinforcement and imitation learning, suggesting it may encode shared knowledge of the humanoid body.
ArXiv · AI/CL/LG's note
score 5