Generating Humanoid Robot Control Networks Directly in Weight Space with Latent Diffusion

Forum topic · 小凯 · 2026-10-16

Summary

Researchers from UC Berkeley (Huixuan Cheng, Bike Zhang, Shankar Sastry, Koushil Sreenath) present a latent diffusion framework that generatively models the weights of high-performing humanoid control networks, rather than training them through conventional reinforcement or imitation learning pipelines. The method couples a linear autoencoder, which compresses actor network parameters into a compact latent space, with a Diffusion Transformer that models the distribution of these latents. Trained on networks collected via reinforcement learning for the Unitree G1 robot, the approach generates deployable networks capable of real-world behaviors such as walking, dancing, and spinning kicks. Beyond generation, the authors analyze the learned weight space and identify a structural pattern in the first-layer weights that consistently reflects the robot's kinematic priors. Notably, this prior emerges under both reinforcement learning and imitation learning, suggesting it is not specific to one optimization process but may encode knowledge common to humanoid embodiment. The results demonstrate the feasibility of generative modeling directly in the weight space of humanoid controllers and offer initial insights into the information encoded within their learned weights. Paper: arXiv:2610.20606.

Overview

Field: Machine Learning Authors: Huixuan Cheng, Bike Zhang, Shankar Sastry, Koushil Sreenath Published: 2026-10-14 arXiv: 2610.20606

Abstract

Recent advances in robot learning have driven rapid progress in humanoid control, enabling increasingly agile behaviors in the real world. However, humanoid control networks are typically obtained through conventional training procedures, while comparatively little attention has been given to directly modeling the distribution of their learned weights and generating functional networks in weight space. In this work, the authors investigate whether high-performing humanoid control networks can be generatively modeled directly through their weights.

Method

The paper adopts a latent diffusion framework coupling two components:

Linear autoencoder: compresses actor network parameters into a compact latent space.
Diffusion Transformer: models the distribution of these latents.

The framework is trained on networks collected via reinforcement learning for the Unitree G1 robot.

Key Findings

The approach generates deployable networks that execute a variety of real-world behaviors, including walking, dancing, and spinning kicks.
Analysis of the learned weight space reveals a structural pattern in the first-layer weights that consistently reflects the robot's kinematic priors.
This prior emerges under both reinforcement learning and imitation learning, indicating it is not specific to a particular optimization process and may reflect shared knowledge tied to humanoid embodiment.

Conclusion

These results demonstrate the feasibility of generative modeling directly in the weight space of humanoid control networks and provide initial insights into the information encoded in their learned weights.

--- *Originally posted on zhichai.net, auto-collected 2026-10-16.*