ActionMap: Robot Policy Learning via Voxel Action Heatmap

Pei Yang^1,*, Hai Ci^1,*, Yanzhe Chen^1,*, Qi Lv¹, Han Cai², and Mike Zheng Shou^1,✉

¹ Show Lab, National University of Singapore ² NVIDIA

ActionMap replaces the single-point action decoder of vision-language-action (VLA) models with a voxel action heatmap, improving success rate, data efficiency, and convergence across LIBERO simulation and real-world Franka manipulation.

ActionMap teaser: single-point paradigm vs. voxel action heatmap paradigm

Many VLAs predict each action as a single point in continuous space, whether through token bins, L1 regression, or flow matching, leaving the geometry of neighboring actions unused. ActionMap instead predicts three voxel heatmaps, over translation, rotation, and gripper, trained by cross-entropy against a soft Gaussian blob centered on the ground-truth action.

ActionMap action head and training objective

ActionMap swaps only the action decoder, so the same head drops into different VLA backbones without changing anything else. The figure below shows one such instantiation, replacing OpenVLA-OFT's L1 regression head, where each heatmap is decoded into a continuous action by hard argmax or top-k soft-argmax.

ActionMap as a drop-in decoder, illustrated on OpenVLA-OFT alongside OpenVLA

On LIBERO, dropping ActionMap into both backbones lifts the four-suite average at matched training steps, by +8.2% over OpenVLA-OFT's L1 head and +1.6% over π_0.5's flow-matching head, while converging at least as fast.

OpenVLA-OFT backbone

π_0.5 backbone

The advantage widens as data shrinks. With only 10% of the demonstrations, ActionMap holds 93.2% on LIBERO-Spatial while the L1 regression head collapses to 67.2%.

OpenVLA-OFT backbone

π_0.5 backbone

On a real Franka Research 3 robot, ActionMap wins on every task at full data and reduces end-effector grasp-position error by two to three times. Below, each task pairs a heatmap rollout with its success counts, with the predicted heatmaps staying sharp and well-localized throughout the episode.

OpenVLA-OFT OpenVLA-OFT + Ours

Pick Brick

Sweep Brick

Insert Card

@article{actionmap,
    title={ActionMap: Robot Policy Learning via Voxel Action Heatmap},
    author={Pei Yang and Hai Ci and Yanzhe Chen and Qi Lv and Han Cai and Mike Zheng Shou},
    year={2026},
    eprint={2606.06904},
    archivePrefix={arXiv},
    primaryClass={cs.RO},
    url={https://arxiv.org/abs/2606.06904},
}