Recoleta
GitHub
Source note

ActionMap: Robot Policy Learning via Voxel Action Heatmap

Authors
Pei Yang, Hai Ci, Yanzhe Chen, Qi Lv, Han Cai, Mike Zheng Shou
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot Policy LearningAction DecoderVoxel HeatmapData Efficiency
Open arXivMarkdown source

Summary

ActionMap replaces single-point action decoders in vision-language-action robot policies with voxel heatmaps over translation, rotation, and gripper actions. It reports higher success rates, faster training, and better low-data performance on LIBERO and real Franka tasks without changing the VLA backbone.

Problem

  • Current VLA action decoders usually output one action point through token bins, L1 regression, or flow matching, so training loses the geometry of nearby actions.
  • This matters for robot manipulation because small end-effector errors can cause grasp, sweep, or insertion failures.
  • The paper tests whether action representation alone can improve VLA policies, separate from larger backbones or more robot data.

Approach

  • ActionMap is a drop-in MLP action head that reads the VLA action-token hidden states and replaces the native decoder.
  • The head predicts separate probability grids for 3D translation, 3D rotation, and the binary gripper command.
  • Training converts each ground-truth action into a soft Gaussian blob over the voxel grid and uses cross-entropy against the predicted heatmap.
  • At inference, it recovers a continuous action with top-k soft argmax, using k=10 and temperature 1.0 in the main experiments; the gripper uses argmax.
  • The paper plugs the head into OpenVLA-OFT and pi0.5 while leaving the rest of each policy unchanged.

Results

  • On LIBERO, ActionMap improves the four-suite average over OpenVLA-OFT L1 regression from 89.1% to 97.3%, a +8.2 point gain at matched 10K training steps.
  • On LIBERO with pi0.5, it improves the four-suite average from 96.9% to 98.5%, a +1.6 point gain at matched 30K training steps.
  • On LIBERO-Long, the largest gains are +26.6 points on OpenVLA-OFT and +4.8 points on pi0.5.
  • At 10% LIBERO-Spatial data, equal to 43 demonstrations, ActionMap gets 93.2% while OpenVLA-OFT L1 regression gets 67.2%, a +26.0 point gap.
  • On real Franka tasks at full data, ActionMap succeeds in 20/30 trials versus 7/30 for OpenVLA-OFT L1 regression; at 50 demonstrations per task, it gets 14/30 versus 4/30.
  • On the Franka Pick task, ActionMap reports 4.8 mm grasp-position error at full data and 15.0 mm at 50 demonstrations, about 2 to 3 times lower error than the L1 regression head.

Link