rkstgr/aic-cheatcode-rollouts-v1

AIC CheatCode Rollouts v1

Twenty trial rollouts of the CheatCode scripted policy on the AIC cable-insertion task, recorded in Rerun .rrd format with AV1-compressed video.

Each file is a self-contained Rerun recording containing synchronized:

• 3× camera streams (left, center, right) — Basler acA2440-20gc at 576×512, AV1-encoded MP4 (AssetVideo + VideoFrameReference).
• TCP pose (position + quaternion), velocity (linear + angular), and 6-dof error.
• Joint positions (7 joints).
• Force/torque wrench (3+3) plus pre-computed |F|.
• Commanded action twist (linear, angular) — note: CheatCode publishes via MotionUpdate.pose, so the velocity-action streams here are all zero. See index.json:t3_message and the scoring file for what the policy did.
• 3D world frame of the TCP point.

Trajectories are trimmed to the motion-relevant window (first non-trivial TCP velocity to the end of the post-success held-pose period).

Contents

rrd/trial_01.rrd … rrd/trial_20.rrd     # 20 trial recordings (~1 MB each)
index.json                              # per-trial metadata + scoring (used by the viewer)

Summary

Trials	20
Successes (T3 = 75)	15 / 20
Plug types	12 SFP, 8 SC
SFP success	11 / 12 (92 %)
SC success	4 / 8 (50 %)
Force-penalized (Tier 2: >20 N for >1 s)	0 / 20
Camera resolution	576 × 512
Codec	AV1 (libsvtav1 preset 10 CRF 50) in MP4
Per-trial size	≈ 1 MB
Trim window	first arm motion → motion-stopped + 1.5 s tail
Recorded	2026-05-15

Per-trial metadata

(All have Tier 1 = 1.0. T3 = 75 indicates a fully-seated cable; lower values mean partial / no insertion. See AIC scoring docs.)

trial	plug	T3	success	trim dur (s)
1	sfp	75.0	✓	43.0
2	sfp	75.0	✓	54.8
3	sfp	75.0	✓	43.9
4	sc	75.0	✓	96.5
5	sfp	75.0	✓	65.7
6	sc	39.5	✗	63.9
7	sfp	75.0	✓	62.7
8	sfp	75.0	✓	66.8
9	sfp	75.0	✓	62.4
10	sc	75.0	✓	82.5
11	sc	75.0	✓	82.5
12	sc	39.7	✗	63.6
13	sfp	75.0	✓	56.8
14	sc	75.0	✓	88.8
15	sfp	75.0	✓	59.2
16	sc	39.7	✗	55.4
17	sfp	75.0	✓	61.9
18	sfp	75.0	✓	41.0
19	sfp	38.6	✗	47.7
20	sc	22.7	✗	61.5

The four SC partials all stop ≈ 1 cm short of full seating — see cheatcode-sc-seating. Trial 19 is the most interesting outlier: a partial SFP, which usually seats cleanly.

How to view

One-file open in Rerun (no setup):

pip install rerun-sdk==0.26.2
rerun --serve rrd/trial_01.rrd

Filterable web viewer with all 20 trials + scoring columns: in the rkstgr/intrinsic-aic repo under tools/trial_viewer/. Clone it, drop these files into tools/trial_viewer/viewer/, and serve.

Generation

These recordings come from record_smoke_runner.sh invoking CheatCode against the AIC eval sim (Gazebo). Pipeline: record → AV1-compress → trim → index. Full source in rkstgr/intrinsic-aic under tools/trial_viewer/pipeline/.

Known limitations

• Wrench is captured untared (raw /fts_broadcaster/wrench, ≈ 7 N baseline from cable/gripper weight); the AIC scoring subtracts fts_tare_offset before applying the 20 N penalty threshold. Use index.json[*].force_penalized (parsed from scoring.yaml) for the canonical answer.
• The action/linear and action/angular streams are all zeros: CheatCode commands MotionUpdate.pose, not .velocity. A future recorder version should capture .pose too.