ActionMap (50 demos per task)

We replaced only OpenVLA-OFT’s L1 regression head with the paper’s voxel heatmap head, and trained both on the same 50 demonstrations per task with the paper’s real-robot recipe. Both heads match the official code. The final round re-evaluates these same models on the station’s blend runtime instead of the synchronous one. Differences from the paper:

• Two arms and three cameras. Each step predicts 14 action values instead of 7, and ActionMap has one heatmap branch per arm on a shared trunk.
• Delta actions on absolute targets. Both heads predict per-step deltas that are added to the measured pose, with normalization ranges fitted per chunk position. Rotations are rotation vectors instead of Euler angles.
• Scarce data. Each task’s 50 training episodes are a fixed, seeded subset of the 276 to 288 available demonstrations, stratified by which arm moves.
• Training budget. Both heads train for 4,000 steps at a constant learning rate of 5e-4 and batch 64, without a validation split.
• Chunk timing. Chunks are 16 steps at 30 Hz, about the paper’s 0.5 s horizon.
• Blend execution. The final round requests each next chunk 800 ms ahead and crossfades into it over 600 ms. Because a chunk lasts only 533 ms, a new request starts as soon as the previous one returns, instead of each chunk executing in full before replanning as OFT does.
• Evaluation. Each task runs 30 placements with one trial each, instead of the paper’s 10 trials.
• Hardware. The pour pair trained on four A100s and the other pairs on four H100s, always the same within a task.