dhruv1672/so101-peg-insertion-eval-results

SO-101 peg insertion: closed-loop evaluation records

Every closed-loop simulator episode behind the reported results of the SmolVLA SO-101 peg-insertion models: 436 episode records from 12 test runs (2026-11-06 to 2026-11-07). Published for transparency and reanalysis. Private.

Models: dhruv1672/smolvla-so101-peg-insertion (revisions main, relative-020000, relative-expertv3-005000, relative-combined-vision-005000) and dhruv1672/smolvla-so101-peg-insertion-pan-shaped (update-10000). Data: dhruv1672/so101-isaac-sim-peg-insertion-task, dhruv1672/so101-isaac-sim-peg-insertion-pan-shaped. Code, scene specs and evaluator: https://github.com/pateldhruv1672/so-101 (tools/eval_grid_vla.py, variations/closed_loop/). The consolidated experiment log with interpretation and caveats is TEST_LOG.md here.

Success labels (per episode)

Every episode carries an explicit success label in labels/episode_outcomes.csv and labels/episode_outcomes.json (one row per episode, keyed by record_index (0-based line number in episodes.jsonl)). Labels are copied from the evaluator's own per-episode record written at test time (source_file column, field success); nothing was inferred from video. Success = peg tip at least 20 mm below the opening and inside the bore within 20 s.

Model labels use the project's canonical IDs (M1-M7 with the serving variant after a plus sign; see docs/MODELS.md in the code repo): none = no serving filter; kalman = Kalman filter on arm targets; speed limit = per-joint speed limit at the expert's 99.5th percentile; replan 5 = execute 5 actions per predicted chunk instead of 10; sampling 20 = 20-candidate medoid sampling per chunk; M2 @ 002500, M5 @ 040000 = intermediate checkpoint of that model's training run. Scripted controllers are references, not learned policies.

Run / experiment	Model label	0.5 mm	0.75 mm	1 mm	1.25 mm	1.5 mm	1.75 mm	2 mm	2.5 mm	Total	Successful episodes
horizon_speed_20261106	Scripted expert (privileged state, reference)					10/10				10/10	0, 1, 2, 3, 4, 5, 6, 7, 8, 9
horizon_speed_20261106	M1 + none					1/10				1/10	12
horizon_speed_20261106	M1 + speed limit					0/10				0/10	none
horizon_speed_20261106	M1 + replan 5					0/10				0/10	none
rel002500_pilot_20261106	Scripted expert (privileged state, reference)					3/3				3/3	40, 41, 42
rel002500_pilot_20261106	M2 @ 002500 + none					0/3				0/3	none
rel002500_pilot_20261106	M2 @ 002500 + speed limit					0/3				0/3	none
rel002500_pilot_20261106	M2 @ 002500 + replan 5					1/3				1/3	50
rel012500_20261106	Scripted expert (privileged state, reference)					1/1				1/1	52
rel020000_vs_abs_20261106	M1 + none					2/10				2/10	54, 62
rel020000_vs_abs_20261106	Scripted expert (privileged state, reference)					10/10				10/10	63, 64, 65, 66, 67, 68, 69, 70, 71, 72
rel020000_vs_abs_20261106	M2 + none					4/10				4/10	78, 79, 81, 82
rel020000_vs_abs_20261106	M2 + replan 5					1/10				1/10	88
seen_unseen_pan1mm_20261107	M5 @ 040000 + none			2/6						2/6	93, 98
seen_unseen_pan1mm_20261107	M2 + kalman			2/6						2/6	99, 102
seen_unseen_pan1mm_kf_20261107	M5 @ 040000 + kalman			1/6						1/6	110
seen_unseen_pan1mm_kf_20261107	M2 + none			1/6						1/6	115
seen_unseen_training16_20261106	Scripted expert (privileged state, reference)	2/4	4/4	4/4	4/4	4/4	4/4	4/4	4/4	30/32	118, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148
seen_unseen_training16_20261106	M2 + none	0/4	2/4	2/4	2/4	0/4	1/4	2/4	1/4	10/32	154, 156, 158, 160, 163, 164, 170, 174, 176, 180
smoke_vla_20261106	M1 + none					0/3				0/3	none
tight_fresh_ft_kf_20261106	Scripted expert (privileged state, reference)	7/10	8/10	10/10						25/30	185, 186, 187, 188, 190, 192, 193, 195, 196, 197, 198, 199, 200, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213
tight_fresh_ft_kf_20261106	M3 + none	1/10	1/10	1/10						3/30	220, 228, 238
tight_fresh_ft_kf_20261106	M3 + kalman	0/10	0/10	1/10						1/30	268
tight_fresh_ft_kf_20261106	M2 + none	4/10	3/10	2/10						9/30	275, 278, 280, 283, 285, 288, 293, 295, 298
tight_fresh_old_kf_20261106	Scripted expert (privileged state, reference)	7/10	6/7							13/17	305, 306, 307, 308, 310, 312, 313, 315, 316, 317, 318, 319, 320
tight_fresh_old_kf_20261106	M2 + none	2/10	0/7							2/17	322, 327
tight_fresh_old_kf_20261106	M2 + kalman	0/10	0/7							0/17	none
tight_fresh_old_kf_20261106	M2 + speed limit	1/10	1/7							2/17	361, 365
vision_sampling15_20261106	Scripted expert (privileged state, reference)	0/1								0/1	none
vision_sampling15_20261106	M2 + sampling 20 + kalman	0/1								0/1	none
vision_sampling15_20261106	M4 + none	0/1								0/1	none
vision_sampling15_20261106	M4 + sampling 20 + kalman	0/1								0/1	none
vision_sampling15_batched_20261107	Scripted expert (privileged state, reference)	2/5	4/5	5/5						11/15	378, 380, 381, 383, 384, 385, 386, 387, 388, 389, 390
vision_sampling15_batched_20261107	M2 + sampling 20 + kalman	1/5	2/5	2/5						5/15	394, 396, 399, 401, 404
vision_sampling15_batched_20261107	M4 + none	1/5	1/5	1/5						3/15	408, 413, 418
vision_sampling15_batched_20261107	M4 + sampling 20 + kalman	1/5	1/5	0/5						2/15	424, 426

Totals: 436 episodes, 155 successful (52 by learned policies, 103 by reference controllers). Counts are within one run only; do not pool runs.

What a record is

Isaac Sim 6.0.1 / Isaac Lab 3.0, visible GUI, four (once eight) independent robots in one window on a shared physics clock, each with its own scene, cameras and controller. Policies compared within a run see the same scenes and prediction-noise seeds. Success follows the assignment: peg tip at least 20 mm inside the bore within 20 simulated seconds. Simulated 30 Hz is the control clock; physics waits for model inference, so it is not wall-clock 30 Hz.

No videos and no continuous camera recordings exist for these model tests. The evaluator saved numbers, not image streams:

File	Contents
episodes.jsonl	one row per episode (viewer table): run, policy, model reference, checkpoint, scene, clearance, success, depths, time, force, slip, final lateral error, peg-tip high-pass motion, largest commanded pan step and pan range
runs/<run>/<policy>/episodes/*.json	full record: assignment fields, diagnostics, every 30 Hz commanded arm target (commanded_arm_deg), peg-tip trajectory at 10 Hz (trajectory), jitter metrics, applied scene variation, model server settings, reset options, code commit
runs/<run>/<policy>/frames/*.png	the initial wrist and overhead frames of each scene only
runs/<run>/<policy>/results.json, runs/<run>/summary.*, manifest.json	the evaluator's own per-policy results and summaries
TEST_LOG.md	protocols, results and caveats for all tests, including ones without records here (offline probes, collections)

Depth: insertion_depth_mm is the assignment's vertical-depth maximum and can be large while the peg is outside the bore. Use success together with in_hole_max_depth_mm.

Absolute local paths were replaced by project-relative paths or <local>; nothing else in the records was changed.

Runs

Run	Episodes	Status	What it tested
horizon_speed_20261106	40	complete	Absolute model at 1.5 mm, 10 scenes: plain, re-plan every 5, speed-limited, expert.
rel002500_pilot_20261106	12	complete, early pilot	Relative model at 2,500 of 20,000 steps, 3 scenes at 1.5 mm.
rel012500_20261106	1	incomplete	Relative model at 12,500 steps; one episode recorded before the run was stopped. Not a result.
rel020000_vs_abs_20261106	40	complete	Relative 020000 vs absolute 020000 at 1.5 mm, 10 matched scenes, plus re-plan-5 and expert.
seen_unseen_pan1mm_20261107	12	complete	1.0 mm, 3 seen + 3 fresh scenes, ramped reset: pan-shaped model at 10k updates (no filter) vs relative 020000 + Kalman.
seen_unseen_pan1mm_kf_20261107	12	complete	Same 6 scenes, seeds and reset: pan-shaped model + Kalman vs relative 020000 without filter.
seen_unseen_training16_20261106	64	complete	Relative 020000 and expert on 16 replayed training scenes and 16 fresh partners, 2 per clearance x 8 clearances.
smoke_vla_20261106	3	complete, plumbing smoke	First closed-loop smoke of the absolute model (one standard, one negation, one colour job); not a performance test.
tight_fresh_ft_kf_20261106	120	complete	30 fresh scenes, 10 each at 0.5/0.75/1.0 mm: frozen-vision expert-v3 continuation (new), the same + Kalman, relative 020000 (old), expert.
tight_fresh_old_kf_20261106	68	stopped by the user after 17 of 30 scenes	Second window on the same 30 scenes: relative 020000 plain / + Kalman / speed-limited, expert. Only 0.5 mm (10) and 0.75 mm (7) completed; the in-flight scene was not scored.
vision_sampling15_20261106	4	interrupted	First sequential attempt of the vision comparison; 4 completed episodes, superseded by the batched run. Not pooled.
vision_sampling15_batched_20261107	60	complete	15 fresh scenes, 5 each at 0.5/0.75/1.0 mm: relative 020000 + sampling + Kalman, combined vision model plain and with sampling + Kalman, expert.

Success by run and policy

Counts are within one run only. Do not pool runs: several reuse the same scenes, and some runs were stopped or interrupted.

Run	Policy	Model	Success
horizon_speed_20261106	expert	scripted expert (privileged state)	10/10
horizon_speed_20261106	vla	dhruv1672/smolvla-so101-peg-insertion@main	1/10
horizon_speed_20261106	vla_lim	dhruv1672/smolvla-so101-peg-insertion@main	0/10
horizon_speed_20261106	vla_n5	dhruv1672/smolvla-so101-peg-insertion@main	0/10
rel002500_pilot_20261106	expert	scripted expert (privileged state)	3/3
rel002500_pilot_20261106	vla	unpublished intermediate checkpoint	0/3
rel002500_pilot_20261106	vla_lim	unpublished intermediate checkpoint	0/3
rel002500_pilot_20261106	vla_n5	unpublished intermediate checkpoint	1/3
rel012500_20261106	expert	scripted expert (privileged state)	1/1
rel020000_vs_abs_20261106	abs	dhruv1672/smolvla-so101-peg-insertion@main	2/10
rel020000_vs_abs_20261106	expert	scripted expert (privileged state)	10/10
rel020000_vs_abs_20261106	vla	dhruv1672/smolvla-so101-peg-insertion@relative-020000	4/10
rel020000_vs_abs_20261106	vla_n5	dhruv1672/smolvla-so101-peg-insertion@relative-020000	1/10
seen_unseen_pan1mm_20261107	new	dhruv1672/smolvla-so101-peg-insertion-pan-shaped@update-10000	2/6
seen_unseen_pan1mm_20261107	old_kf	dhruv1672/smolvla-so101-peg-insertion@relative-020000	2/6
seen_unseen_pan1mm_kf_20261107	new_kf	dhruv1672/smolvla-so101-peg-insertion-pan-shaped@update-10000	1/6
seen_unseen_pan1mm_kf_20261107	old	dhruv1672/smolvla-so101-peg-insertion@relative-020000	1/6
seen_unseen_training16_20261106	expert	scripted expert (privileged state)	30/32
seen_unseen_training16_20261106	vla	dhruv1672/smolvla-so101-peg-insertion@relative-020000	10/32
smoke_vla_20261106	vla	dhruv1672/smolvla-so101-peg-insertion@main	0/3
tight_fresh_ft_kf_20261106	expert	scripted expert (privileged state)	25/30
tight_fresh_ft_kf_20261106	new	dhruv1672/smolvla-so101-peg-insertion@relative-expertv3-005000	3/30
tight_fresh_ft_kf_20261106	new_kf	dhruv1672/smolvla-so101-peg-insertion@relative-expertv3-005000	1/30
tight_fresh_ft_kf_20261106	old	dhruv1672/smolvla-so101-peg-insertion@relative-020000	9/30
tight_fresh_old_kf_20261106	expert	scripted expert (privileged state)	13/17
tight_fresh_old_kf_20261106	old	dhruv1672/smolvla-so101-peg-insertion@relative-020000	2/17
tight_fresh_old_kf_20261106	old_kf	dhruv1672/smolvla-so101-peg-insertion@relative-020000	0/17
tight_fresh_old_kf_20261106	old_lim	dhruv1672/smolvla-so101-peg-insertion@relative-020000	2/17
vision_sampling15_20261106	expert	scripted expert (privileged state)	0/1
vision_sampling15_20261106	old_s20_kf	dhruv1672/smolvla-so101-peg-insertion@relative-020000	0/1
vision_sampling15_20261106	vision	dhruv1672/smolvla-so101-peg-insertion@relative-combined-vision-005000	0/1
vision_sampling15_20261106	vision_s20_kf	dhruv1672/smolvla-so101-peg-insertion@relative-combined-vision-005000	0/1
vision_sampling15_batched_20261107	expert	scripted expert (privileged state)	11/15
vision_sampling15_batched_20261107	old_s20_kf	dhruv1672/smolvla-so101-peg-insertion@relative-020000	5/15
vision_sampling15_batched_20261107	vision	dhruv1672/smolvla-so101-peg-insertion@relative-combined-vision-005000	3/15
vision_sampling15_batched_20261107	vision_s20_kf	dhruv1672/smolvla-so101-peg-insertion@relative-combined-vision-005000	2/15

Model references

Model reference	Episodes
dhruv1672/smolvla-so101-peg-insertion-pan-shaped@update-10000	12
dhruv1672/smolvla-so101-peg-insertion@main	43
dhruv1672/smolvla-so101-peg-insertion@relative-020000	161
dhruv1672/smolvla-so101-peg-insertion@relative-combined-vision-005000	32
dhruv1672/smolvla-so101-peg-insertion@relative-expertv3-005000	60
scripted expert (privileged state)	119
unpublished intermediate checkpoint	9

policy suffixes: _kf constant-velocity Kalman filter on the arm targets; _lim per-joint speed limit at the expert's 99.5th percentile; _n5 re-plan every 5 actions instead of 10; _s20 20-candidate medoid sampling per chunk. expert is the privileged-state scripted controller that produced the training data, not a learned policy. Policy labels such as new, old and vision are local to their run; the model reference column resolves them.