arrow-hf's picture
move_can_pot 100005 real mask tracking (no occlusion in sim; can stays 371-2237px)
3a48626 verified
|
Raw
History Blame Contribute Delete
1.12 kB

move_can_pot seed 100005 — real mask tracking (poselib + OWLv2)

move_can_pot_100005_mask.mp4 — every frame's SAM2 mask overlay for the actual sim episode (mask_backend=groundingdino, detector=owlv2, occlusion recovery ON). magenta = can (obj1), cyan = pot (obj2); the number is each mask's pixel area.

Finding: NO occlusion failure in this run

The can mask stays correctly on the pink can the WHOLE episode (area 371-2237 px, never empty). It dips to ~371 px during the grasp (gripper partially covers the can) but never loses lock; pot stays tracked throughout. Episode SUCCEEDs. Occlusion recovery fired 0 times — there is nothing to recover here.

Sample frames: f0001 (start, can=1724), f0049 (grasp, can=703 partial), f0066 (end, can=1758, held beside the pot).

So in sim, move_can_pot 100005 does not reproduce a mask-loss occlusion. The recovery mechanism itself is verified separately (controlled black-out: area 3490 -> 0 -> recovered 3490). The mask-tracking failure reported on the real robot comes from real appearance/position drift during occlusion, which clean sim frames here don't produce.