Action Map Policy turns robot actions into pixel classification, beating Diffusion Policy by 50-70% on real-world tasks

机器之心 · wechat · 2026-09-09

PhD student Haojie Huang (Northeastern University, advised by Robert Platt and Robin Walters) introduces Action Map Policy (AMP), a new action representation for robot manipulation: 3D motion is invertibly converted to 3D keypoint trajectories, projected onto camera planes as 2D pixel trajectories, and prediction becomes pixel-level classification via cross-entropy over keypoint probability maps.

This unifies vision and action in image space (analogous to token space in LLMs), naturally models multimodal action distributions, and outputs the full distribution in a single forward pass. It achieves 1mm positional and 1.3° rotational 3D accuracy at 224×224 resolution, recovering 3D trajectories via multi-view triangulation.

Key results: in a fine-grained laser-pointer test with only 40 demos, AMP hit 100% success while Diffusion Policy showed mode collapse. On three real-world breakfast tasks (coffee, toast, steamed egg), AMP outperformed DiffPo and ACT by 50-70%; inference takes 13.80ms versus 93.53ms for 16-step DDIM DiffPo, suiting real-time closed-loop control.

Original post →

More from Embodied

Embodied channel →