CUPID

We fixed the scoring bug that withdrew the previous round, rescored the demonstrations of all three tasks (pouring screws, placing boxes and clamping a component) and retrained only the CUPID arm. CUPID scores every demonstration from 30 rollouts of a base Diffusion Policy trained on all demonstrations and keeps the top-scoring half. The control keeps a random half. The base and random arms and their evals were reused from the previous round, and all arms train the same platform Diffusion Policy and are evaluated on the same 30 placements. Differences from the paper:

• Corrected scorer. The scores now use the single inverse of the gradient covariance. They were recomputed from the saved feature sums without new rollouts, and we checked that the retrained models trained on the corrected halves.
• Same placements. CUPID scored the same 30 placements later used for the final evals, which favours CUPID.
• Natural data. Our datasets hold naturally collected excellent and ok demonstrations (60 to 92 percent excellent), not the paper’s deliberately mixed-quality sets, which leaves less to filter out.
• Floor on clamping. The base clamping policy succeeded on 1 of 30 rollouts, so its scores came almost entirely from failures.
• Approximations for local compute. Training frames were sampled every 5, 15 or 20 frames depending on the task, and a sparse random projection replaced TRAK’s dense one. Both add noise but no bias.
• Platform policy. All arms use the platform’s three-camera Diffusion Policy with its overlapping-chunk runtime instead of the paper’s implementation.