CUPID

This round’s CUPID scores were wrong, and the round is withdrawn. A bug in our scorer applied the inverse of the gradient covariance twice, so the scores used its squared inverse instead of the method’s single inverse and did not test CUPID. We had trained a base Diffusion Policy on all demonstrations of each of three tasks (pouring screws, placing boxes and clamping a component), evaluated it on 30 placements, scored every demonstration from those rollouts and kept the top-scoring half. The control kept a random half, and both arms trained the same platform Diffusion Policy and were evaluated on the same 30 placements. Differences from the paper:

• Wrong scores. The scorer inverted an already accumulated Gram matrix a second time, so the scores used its squared inverse. The kept halves and the CUPID results are invalid.
• 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.