AMP reframes robot manipulation as pixel classification to dodge action-space explosion

jiqizhixin · x · 2026-09-18

LLMs showed cross-entropy prediction scales well on 1D sequences, but robot actions are high-dimensional: 6–7 DOF per timestep, so even coarse discretization (10 bins per dimension) yields 1 million possible joint actions, and finer discretization or longer horizons make the space explode — so the cross-entropy paradigm behind LLMs can't directly extend to robot learning.

Action Map Policy (AMP) proposes a new paradigm for robot manipulation learning: it reframes high-precision 3D closed-loop manipulation as pixel classification in image space. Instead of predicting action tokens in a combinatorially exploding space, the model classifies where to act on an action map — a 2D representation encoding 3D motion as spatial positions — bringing cross-entropy prediction from language into robotics without the dimensionality blowup.

Original post →

More from Embodied

Embodied channel →