ActionMap (full data)

We replaced only OpenVLA-OFT’s L1 regression head with the paper’s voxel heatmap head, and trained both on each task’s full demonstration set with the paper’s real-robot recipe. Because OFT’s custom transformers fork does not run on our stack, we rebuilt OpenVLA-7B on a newer transformers release; both heads follow the official code. This round evaluates them on the station’s synchronous runtime. 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 between consecutive commands, as in the paper, and integrate them from the measured pose. The first step is taken from the measured pose and normalized separately, because this robot’s commands lead its measured pose by up to 20 cm.
• Training budget. Both heads train for 4,000 steps at a constant learning rate of 5e-4 and batch 64, without a validation split. On clamp, whose episodes are longest, this covers only 0.8 passes over the data.
• Chunk timing and execution. Chunks are 32 steps at 30 Hz, about 1.07 s, and each chunk executes in full before the next request, as in OFT.
• Evaluation. Each task runs 30 placements with one trial each, instead of the paper’s 10 trials. One pour baseline placement ran twice, and the paired comparison counts it as half a success.
• Hardware. All six models trained on four A100 40 GB GPUs, with two micro-batches of 8 per GPU.