Bimanual Garment Folding: Research in Progress
A folding-policy port and camera-health investigation. Strict evaluation has not shown an adaptation gain.
Problem
A folding rollout can look plausible while its policy receives stale camera frames. The experiment needs fresh observations, held-out garments, and an external fold checker. A demonstration replay or a saved success label alone is not enough evidence.
Approach
Port the policy and observation path into Isaac Sim, then grade closed-loop rollouts with the challenge's garment-fold checker. The latest evaluation checks camera freshness and treats rendering failures as experiment failures. Compare the baseline and adapted checkpoints on matching short-pants poses; retain incomplete runs instead of converting them into successful results. This is a research implementation of an external method, not a verified reproduction of the original challenge performance.
What I built
Built the Isaac Sim integration, observation-path diagnostics, evaluation adapters, and Slurm experiment workflow. Investigated stale-image failures and preserved both task failures and infrastructure failures in the results. The original policy method and challenge checker belong to their authors.
Result
Strict short-pants evaluation: baseline 8/24, adapted seed 1 3/24; no improvement. Both score 0/4 on unseen garments. Only 5/12 full class/checkpoint evaluation cells completed; the remainder failed or timed out. A separate new seed-0 top rollout failed after 600 actions with 601 validated renders.
What changed the interpretation
An earlier evaluator continued after camera-update errors, so its counts cannot establish visual closed-loop control. New runs check frame freshness; historical clips remain separate. Scores are official checker first-hit successes, not independently verified stable terminal folds.
New failure, with both arm cameras
The adapted seed-0 policy ran 600 actions on a short-sleeve top. All 601 camera renders passed validation and the cloth moved, but the fold checker never passed. These top and wrist views show that one episode at half simulation speed, with a one-second end hold.
Next evaluation gates
Complete the remaining class/checkpoint cells with verified live camera updates, compare matching poses and seeds, and preserve per-episode checker results. Demonstration replay establishes only scene solvability; it is not policy success. No general garment-folding or challenge-winning result is claimed.