move_pillbottle_pad 100004 MPC after bbox fix (raw-pose bbox + lwh EMA) — before/after
Browse files- single_seed/move_pillbottle_pad_100004_mpc_bboxfix/README.md +29 -0
- single_seed/move_pillbottle_pad_100004_mpc_bboxfix/after_fixed_bbox.png +3 -0
- single_seed/move_pillbottle_pad_100004_mpc_bboxfix/after_lastframe.png +3 -0
- single_seed/move_pillbottle_pad_100004_mpc_bboxfix/before_tilted_bbox.png +3 -0
- single_seed/move_pillbottle_pad_100004_mpc_bboxfix/ep0_seed100004_mpc_bboxfix_FAIL.mp4 +3 -0
single_seed/move_pillbottle_pad_100004_mpc_bboxfix/README.md
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# move_pillbottle_pad seed 100004 — MPC after the bbox fix
|
| 2 |
+
|
| 3 |
+
## The bug (confirmed in code)
|
| 4 |
+
poselib output `bbox_size_lwh_m` (ScaleNet) together with the RAW estimator pose,
|
| 5 |
+
so the lwh axes are bound to the raw pose's object frame. But the orchestrator
|
| 6 |
+
reconstructed the 3D bbox corners using the FUSED pose (poses = fuser output).
|
| 7 |
+
When the fuser / rotation-stability changed the rotation (fused != raw — common
|
| 8 |
+
for flat/near-symmetric objects like a pad), the box was mis-oriented and the
|
| 9 |
+
distance constraint (which uses these corners) was computed on a wrong box.
|
| 10 |
+
|
| 11 |
+
## The fix (orchestrator.py)
|
| 12 |
+
1. Reconstruct the bbox with the RAW pose (`obs_poses`, captured per object in
|
| 13 |
+
both batch and serial paths) instead of the fused pose.
|
| 14 |
+
2. Lightly smooth lwh over time (per-object EMA, alpha=0.5) to suppress per-frame
|
| 15 |
+
size jitter (object size is ~constant).
|
| 16 |
+
|
| 17 |
+
## Effect on this seed (still FAIL, but honest)
|
| 18 |
+
| | before (fused-pose bbox) | after (raw-pose bbox + lwh EMA) |
|
| 19 |
+
|---|---|---|
|
| 20 |
+
| reached step | 0→1→2→3 (place) | 0→1→2 (transport) |
|
| 21 |
+
| final distance | -1.421 (mis-oriented-box artifact) | -0.594 (real) |
|
| 22 |
+
| bbox | clearly tilted (before_tilted_bbox.png) | corrected (after_fixed_bbox.png) |
|
| 23 |
+
|
| 24 |
+
Before, the tilted box *spuriously* satisfied the transport constraint and the
|
| 25 |
+
plan jumped to place. After the fix the box is accurate and correctly shows the
|
| 26 |
+
transport is incomplete: the gripper lifts the pillbottle but never brings it
|
| 27 |
+
down onto the pad (after_lastframe.png: holding=+1.0, distance=-0.594). So this
|
| 28 |
+
seed is a genuine control/policy failure — the fix removed the bbox artifact that
|
| 29 |
+
was masking it.
|
single_seed/move_pillbottle_pad_100004_mpc_bboxfix/after_fixed_bbox.png
ADDED
|
Git LFS Details
|
single_seed/move_pillbottle_pad_100004_mpc_bboxfix/after_lastframe.png
ADDED
|
Git LFS Details
|
single_seed/move_pillbottle_pad_100004_mpc_bboxfix/before_tilted_bbox.png
ADDED
|
Git LFS Details
|
single_seed/move_pillbottle_pad_100004_mpc_bboxfix/ep0_seed100004_mpc_bboxfix_FAIL.mp4
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:03ad5cc87a3a56f94a8f4528eeedd3e6a66fc10784678af7c7a800eafa00c687
|
| 3 |
+
size 14980246
|