arrow-hf's picture
mpc_fail_rerun_400: combined top-level summary README
2a6bafe verified
|
Raw
History Blame Contribute Delete
1.19 kB

MPC failures re-run at max_steps 400 (were FAIL at 200)

Took every seed where MPC failed (env_success=False) at max_steps=200 and re-ran it at max_steps=400 — to separate "ran out of time" from "genuinely can't do it". Same v3-valid config per task (poselib + MPC; bread=economicgrasp, canpot=object_center; completion_frames=2; fling detection on).

task failed @200 recovered @400 still FAIL flung @400 timeout
move_can_pot 12 1 (8%) 10 0 1
place_bread_skillet 25 8 (32%) 6 10 1

Takeaways

  • canpot: doubling the horizon barely helps (1/12). The MPC failures are genuine grasp/place-capability failures, not a step-budget issue.
  • bread: ~1/3 recover with more time, but the extra horizon also lets the VLA fling the bread far more (10/25 = 40%), offsetting much of the gain.
  • Overall: 200 steps is essentially enough for these two tasks; the bottleneck is policy capability (grasp / place / not ejecting the object), not time.

Each task folder: videos/ (re-run episodes), flung_videos/ (ejected; 15-tick grace), results.tsv, failed_at_200_seeds.txt (the input seeds).