Ziploc slide — physics verification on real footage

The gate refused. Easiest tracking case. 0 of 35 calibration cells pass.

Same gate, same gripper-release gating. Pouch is 12/12 tracked. Detection 1.000 on injected. But clean false-positives at 1.000 because the support plane assumption doesn't match reality. Calibration sweep (metres-per-pixel × floor margin, 35 cells) passes 0/35. Not a near-miss on tuning — the support plane assumption itself fails. The failed assumption: a single assumed diameter for a flexible pouch that changes shape during the slide — an input-property gap, not a detector bug. A real deployment's metadata (declared object size, calibrated support plane, IMU, depth) would supply what a public dataset cannot.

← Back to Lab index

What we tested

Does the gate pass on the easiest tracking case?

24/24

injected violations caught · 1.000 · target ≥ 0.950

12/12

clean clips false-flagged · 1.000 · target ≤ 0.020

[0.757, 1.000]

Wilson 95% CI on false positives

Charts

Flag rate by violation class (Ziploc)

flag rate

Methodology

Gripper-release gate on slid pouch. Same gripper-channel Otsu gate. Pouch is 12/12 tracked — easiest tracking case so far.
Calibration sweep. Sweep metres-per-pixel (±2×) and floor margin (0.75–0.95). 35 cells. 0 pass.
Picker funnel. Object found in all 12 clips. Not a tracking failure.

Seeds: 12 clips, single camera, tracked

FP ≤ 0.02
detection ≥ 0.95
coverage complete

Notes

Real tracked footage, gripper-release gated. Dataset: lerobot/aloha_static_ziploc_slide (v1.0). Hardware: ALOHA stationary bimanual (2× ViperX 300), 3× RealSense D435i @ 640×480, 30 FPS. Object: Ziploc bag, ~50 g, polyethylene film, partially filled. Task: Slide, pick, transfer, place. 50 episodes. Flexible pouch changes shape — single assumed diameter is wrong for part of every clip. 0/35 calibration cells pass. Notebook: haga-ziploc-slide-verification. HuggingFace: https://huggingface.co/datasets/lerobot/aloha_static_ziploc_slide

Run it yourself

The notebook below is the production pipeline, inlined and public. It pulls its own data and attaches no private dataset, so you can fork it and run it end to end on free compute. Thresholds are declared before the first episode runs, and a FAIL is published as measured.

Ziploc slide — physics verification on real footageNotebook runlerobot/aloha_static_ziploc_slide

Real robot footage, gripper-release gated. Detection 1.000, FP 1.000. Flexible pouch violates rigid-body assumption; 0/35 calibration cells pass.

Simulation for the stress runs; real tracked robot footage for the physics verification runs. Neither has run on hardware.