arrow-hf/vla-framework-holding-empty-close

vla_framework — empty-close ("抓空") detection investigation

Does the gripper empty-close grasp-failure flag misfire on every grasp, or only on real failures? Investigated on move_pillbottle_pad (SmolVLA, pose_source=robotwin_gt, MPC corrector, chunk_exec=50).

TL;DR

On SmolVLA move_pillbottle_pad the empty-close detector has zero misfires. Every grasp is bimodal: a real grasp lands ~40 N on both jaws (holding confirmed, empty=0), a miss lands 0.0 N (gripper closes on air, empty=1, correct). There is no "held-but-judged-empty" middle case. The grasp-instant red [C] holding people see is holding not yet confirmed (needs a firm two-jaw ≥2 N contact) — a separate signal from the empty flag, and the MPC then engages via REPLAN (phase stall), not via empty.

Exact setup (for config alignment)

• Repo commit: b524368 (feat(robotwin): optional empty-close detection, default OFF)
  • chain: 51946ea fingers-met gate → 54802e8 MPC-engage logging → b524368 toggle
• Env: conda vla_framework; MUJOCO_GL=egl, HF_HUB_OFFLINE=1, CUDA_VISIBLE_DEVICES=1
• Config: config_used/vla_framework_eval.template.yaml (unmodified template)
• Command:

  python scripts/eval_robotwin.py \
    --config config/vla_framework_eval.template.yaml --task move_pillbottle_pad \
    --seed <SEED> --num_episodes 1 --max_steps 200 \
    --model_type smolvla --checkpoint weights/smolvla/robotwin_move_pillbottle_pad_50ep_multi \
    --chunk_exec 50 --no_poselib --contact_trace [--empty_close_detection]
  

• empty_close_detection defaults OFF (b524368); pass --empty_close_detection to re-enable.

Detection logic (RoboTwin adapter, _derive_grasp_flags)

• commanded_closed = gripper target ≤ 0.05
• contact_grasp = both jaw fingers bear ≥ 2.0 N on the same object
• raw_holding = contact_grasp & commanded_closed
• raw_empty_kin = commanded_closed & (finger_opening ≤ 0.05) (fingers met)
• raw_empty = ~raw_holding & raw_empty_kin ← fingers-met gate (fix 51946ea)

Both debounced 2 ticks. empty_close feeds the holding predicate's −1.0 empty penalty. The b524368 toggle zeroes the consumed empty_close (raw signals stay in the trace).

15-seed result — perfectly bimodal, no misfire

outcome	seeds	contact force	held ticks	empty ticks	MPC engage
real grasp (SUCCESS)	100000, 100004, 100006, 100008, 100010	39–46 N	67–78	0	0
miss (FAIL)	100001/2/3/5/7/9/11	0.0 N	0	~50	1 (REPLAN)

• seed 100000 (default, empty OFF): closes @tick 60 with 42 N, holding confirms @tick 61, step advances (NEXT), empty=0, MPC never engages — see videos/pad_seed100000_* (frame 62: [S] holding +1.000, green).
• seed 100001/100002: gripper closes with 0.0 N, raised empty — see videos/pad_seed100002_frame85_empty_gripper.png. empty=1 is correct.

The original held-object false positive (place_bread_skillet seed 100000 → 226 spurious CORRECTs on a held skillet, SUCCESS→FAIL) was the real bug, fixed by the fingers-met gate (51946ea).

pi0 + MPC — the genuine "held but unconfirmed" reproduction

SmolVLA grasps land ~40 N so holding always confirms. pi0 on place_container_plate (seed 100017, pi0_fk_mpc) does NOT: videos/pi0_seed100017_* shows the gripper grasping and lifting/carrying the container (the container bbox tracks the EEF) while [C] holding stays at −0.001 (the unconfirmed placeholder _HOLDING_UNCONFIRMED_MARGIN, red) — the grasp step never advances and the MPC keeps engaging.

Note this is not the empty flag: the container holds the fingers open, so empty_kin = 0 and the empty toggle is irrelevant. It is contact_grasp = 0 — the detector requires both jaws ≥ 2 N on the same body, but pi0 grips the thin container rim lightly / one-sided, so the grasp is never confirmed.

Observed model/corrector dependence (same case, seed 100017):

setup	grasp-time MPC intervention / detection anomaly
pi0 + MPC	yes — holding never confirms, MPC engages at grasp
pi0 + passthrough	no anomaly
smolvla + MPC	no (firm ~40 N grasp confirms)

That pi0+passthrough is clean but pi0+MPC is not points at the MPC corrector perturbing the grasp/gripper in the FK→ee path (under investigation), on top of the contact-confirmation (2 N, both-fingers) being too strict for pi0's light rim grasp. Likely fix directions: lower contact_force_threshold, relax require_both_fingers, and check the corrector isn't smoothing the gripper dim.

Contents

• config_used/ — the exact config used (align yours against this)
• videos/pad_seed100000_success_grasp.mp4 + frame62_holding_confirmed.png — clean grasp, holding green
• videos/pad_seed10000{1,2}_fail_grasp.mp4 + frame85_empty_gripper.png — real misses
• contact_traces/seed*.csv — per-tick contact force / grasp / holding / empty