Datasets:

dhruv1672
/

so101-isaac-sim-peg-insertion-task

Simulated demonstrations of an SO-101 arm inserting a held 20 mm PLA peg into a blind hole, recorded in NVIDIA Isaac Sim 6.0.1 / Isaac Lab and stored in LeRobot v3.0 format. Built to fine-tune lerobot/smolvla_base (revision d9f33c94a60fb382c90dea2164c96845bd955e28).

This is simulation data, not real-robot data. No license has been assigned yet: until one is added, no reuse rights are granted beyond viewing the repository.

Versions

Revision	Episodes	Frames	Contents
`v3.0` (current, `main`)	687	249,892	the 533 original episodes + 154 tight-clearance expert episodes (2026-11-07)
`v3.0-533ep`	533	187,270	the original release (2026-11-06), unchanged

LeRobot loads the v3.0 tag by default. To reproduce models trained on the original release, pass revision="v3.0-533ep".

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("dhruv1672/so101-isaac-sim-peg-insertion-task")                         # 687 episodes
ds_533 = LeRobotDataset("dhruv1672/so101-isaac-sim-peg-insertion-task", revision="v3.0-533ep")  # original


Episodes	687 successful demonstrations: 533 from the original collection (episodes 0-532), 154 tight-clearance expert episodes (533-686)
Frames	249,892 at 30 fps (about 139 min)
Task	`Insert the peg into the hole.` (single task)
Robot type	`so101_follower`
Videos	AV1, CRF 30, 640x480, two cameras, concatenated into ~200 MB files (LeRobot v3 layout)

combined_manifest.json maps every episode to its source collection, raw episode name, clearance and, for the tight-clearance part, the controller, scene id and pair index. conversion_manifest.json is the original release's per-episode manifest for episodes 0-532.

Clearance (mm)	0.5	0.75	1.0	1.25	1.5	1.75	2.0	2.5
Original collection	25	35	86	72	76	92	95	52
Tight-clearance experts	32	58	64
Total	57	93	150	72	76	92	95	52

Features

Key	Shape	Meaning
`observation.state`	6 float32	measured joints: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll (degrees), gripper (0-100)
`action`	6 float32	absolute joint targets, same order and units, held until the next 30 Hz step
`observation.images.wrist`	480x640x3	pinhole wrist camera at the SO-ARM100 MJCF pose, fx = fy = 320
`observation.images.overhead`	480x640x3	fixed camera at (0.22, 0, 0.45) m looking down, fx = fy = 457

Images are recorded before the action of the same step. When fine-tuning smolvla_base, map overhead to camera1 and wrist to camera2 (with one empty camera), which matches its pretraining view names.

Scene and task

SO-101 USD from the Isaac Sim 6.0 asset library (Isaac/Robots/RobotStudio/so101_new_calib), stock joint limits, actuators and gains; base at the origin on the table top, +x forward.
Peg: cylinder 20 mm diameter x 50 mm, PLA density 1240 kg/m3, physically held by the gripper (command 0) from the start.
Hole block 50 x 50 x 30 mm with a 25 mm deep blind hole, no chamfer, diameter 20 mm + 2 x radial clearance.
Friction peg-hole 0.5 (randomized below), peg-gripper 0.8. PhysX TGS at 600 Hz, 30 Hz control.
Start: peg tip 15 mm above the opening and 30 mm closer to the arm than the hole axis, gripper pointing down.
Success: peg tip at least 20 mm below the opening and inside the hole, within 20 s.

How it was collected

Original collection (episodes 0-532). A scripted, privileged expert (true peg and hole poses): slow min-jerk transit, filtered tip correction, descent at 2.5-5 mm/s, the Archimedean spiral search (0.5 mm pitch, 3 mm radius, 5 mm/s) when the tip stalls on the rim, jam back-off, and insertion to 23 mm. Four independent robots ran in one Isaac Lab scene on a shared physics clock.

Tight-clearance experts (episodes 533-686). 120 new randomized scenes, 40 per clearance (0.5, 0.75, 1.0 mm). Every scene was run by two privileged controllers: a filtered expert and a slower extra-slow expert, eight robots in one Isaac Lab window. Both runs of a scene share all randomization, so the pairs show two speeds on the same scene. Table textures are always randomized; 30% of scenes add correlated executed-action noise (shared within a pair).

Per-episode domain randomization (both parts):

Factor	Range
Radial clearance	0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0, 2.5 mm (tight part: 0.5, 0.75, 1.0)
Hole destination	reachable workspace, x 165-228 mm, y +-95 mm, plus +-5 mm hole offset
Peg-hole friction	0.3-0.8
Peg mass	x0.8-1.25
Peg pose in gripper	dx 0-0.5 mm, dy +-1 mm, dz -4-0 mm, tilt +-3 deg
Action latency	0-66 ms
Lighting	intensity x0.4-1.8, azimuth 0-360 deg, elevation 25-85 deg, 3000-7500 K (shared by each simultaneous batch)
Textures	table, peg and block from a 16-texture pool
Camera pose	+-6 mm, +-3 deg per camera
Distractors	0-4 objects, kept clear of the hole and the arm's path

Episodes with executed-action noise still record the clean expert command as action, so the data shows recoveries from small drifts without teaching the noise.

Expert quality (how often the demonstrator itself succeeded)

Only successful episodes are included, so these rates describe the demonstrator, not the data.

Original collection, 600 attempts, 561 successes:

Clearance (mm)	0.5	0.75	1.0	1.25	1.5	1.75	2.0	2.5
Expert success	24/40	32/38	88/95	74/81	76/77	95/95	92/93	52/53

Tight-clearance collection, 240 recordings (120 scenes x 2 controllers), 174 successes:

Clearance (mm)	Filtered expert	Extra-slow expert
0.5	22/40	15/40
0.75	35/40	30/40
1.0	38/40	34/40

The expert is well below 99% reliability at 0.5 mm, so the tightest clearance covers the cases it could solve.

Selection and splits

Not included: 105 failed recordings (39 original, 66 tight-clearance; kept privately for failure analysis) and 48 successful validation episodes (28 original: every 20th batch of four robots; 20 tight-clearance: whole collection batches, both controllers of a scene on the same side). Held-out scenes never appear in training episodes. Raw recordings are not published.

Quality checks

All raw episodes passed an audit (shapes, finite values, frame counts, 30 Hz timestamps, and for the tight part the pairing and split metadata): 0 errors. For the combined dataset: episode and frame totals match both sources; state and action are bit-identical to the source datasets; video files are byte copies (no re-encoding); 88 decoded video frames at episode and file boundaries match the sources exactly; 50-step action chunks are padded at episode ends and never cross into the next episode. The original release's decoded frames match the source JPEGs at 37.3-41 dB PSNR.

Measured model results on this task (for context)

Policies fine-tuned on this data are far from the expert at tight clearances. On 30 fresh scenes (10 each at 0.5, 0.75, 1.0 mm), SmolVLA with relative actions trained on the original 533 episodes inserted 9/30; the scripted expert 25/30. Continuing it on only the 154 tight-clearance episodes (frozen vision) dropped to 3/30. Details: https://huggingface.co/dhruv1672/smolvla-so101-peg-insertion

Limitations

Simulated physics and rendering; contact behaviour at sub-millimetre clearances depends on PhysX settings.
Tight clearances are covered by the episodes the expert could solve; the expert's 0.5 mm success rate is 38-60%.
The two tight-clearance controllers move at different speeds on the same scenes, so a policy sees two styles.
One task string; no language variation. No scripted recovery demonstrations from large disturbances.
Frames were stored as JPEG (quality 92) during collection, then encoded to AV1.
The privileged expert used true object poses that are not in the observations.
The spiral-search stall trigger measures progress over up to 1 s of history rather than exactly 0.3 s.