CUPID

We implemented CUPID’s performance-influence scoring on two tasks, spraying and screw sorting, each with a fixed set of 300 demonstrations. A base policy trained on all 300 demonstrations was evaluated on 30 placements, and CUPID scored every demonstration from those rollouts and kept the top-scoring 150. The control keeps a random 150, and all three arms train the same platform Diffusion Policy and are evaluated on the same 30 placements. Differences from the paper:

• Same placements. CUPID scored the same 30 placements later used for the final evals, which favours CUPID.
• One-sided scores. The base spray policy succeeded on all 30 rollouts and the base screw policy on none, so spray was scored on successes only and screws on failures only, without the contrast between the two that the method relies on.
• Natural data. The datasets hold naturally collected demonstrations, about 80 to 90 percent rated excellent and the rest ok, not the paper’s deliberately mixed-quality sets, which leaves less to filter out.
• Two robots per dataset. Each dataset mixes demonstrations from two robots, and on spray CUPID kept more demonstrations from the evaluation station’s robot than random did (100 against 71 of 150).
• Floor on screws. No arm completed the screw task, so we also report screws placed per episode.
• Dropped task. A third task, putting tools in a toolbox, was dropped before CUPID scoring because its base and random policies both scored 0 of 30.
• Approximations for local compute. Training frames were sampled on a stride, 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.
• Corrected scorer. An earlier round used a squared inverse in its scores and is withdrawn; this round uses the fixed scorer.