Action Head Decoding With Diffusion And L1 Regression

Generalist robot policy model from Open X-Embodiment.

Unique: Implements pluggable action heads (diffusion-based and L1 regression) that decode transformer representations into actions, with diffusion heads enabling multimodal action distributions and uncertainty quantification through iterative denoising. The modular design allows switching between action head types without retraining the transformer.

vs others: Diffusion-based action heads provide better uncertainty quantification and multimodal action support than simple regression heads, while L1 heads offer lower latency for real-time control. The pluggable architecture enables task-specific action head selection without architectural changes.