Notice:

The Chain-of-Action repository has been transferred to ByteDance-Seed/Chain-of-Action.

---

Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation

• 📄 Paper on arXiv
• 🌐 Project Page

🔁 What is Chain-of-Action?

Chain-of-Action rethinks visuomotor policy learning by modeling action trajectories from goal to start, rather than predicting forward from the current state.

Unlike forward policies like ACT or Diffusion Policy, our model starts from the final gripper pose and autoregressively reasons backward—
step by step, toward the current observation.

✅ No extra parameters
✅ No extra data
✅ Just a smarter modeling paradigm

This shift enables strong spatial generalization without tricks.