Policy Learning
Mapping observations to actions — visuomotor policies for robot control.
What Is Policy Learning?
A policy is a function that maps observations (images, state) to actions (joint commands, gripper). Policy learning trains this mapping from data (imitation) or reward (RL). Visuomotor policies use vision as the primary input.
Key Architectures
- ACT (Action Chunking with Transformers) — Predicts action chunks; smooth execution.
- Diffusion Policy — Denoising diffusion for multimodal action distributions.
- Behavior Cloning — Simple supervised learning from demos.
- VLA — Vision-language-action models (OpenVLA, RT-2) with language conditioning.
Related Resources
- Policy Models — ACT, Diffusion Policy, OpenVLA, Octo
- Imitation Learning
- Data Services — Data for training your policy