knpt/yam_bimanual_tool_rack

Dataset Card for YAM bimanual — tool-rack placement

Preview

Egocentric camera view — bimanual robot picking a screwdriver off the bench and inserting it into the tool rack.

Interactive visualizer: huggingface.co/spaces/lerobot/visualize_dataset

66 teleoperated episodes of a bimanual robot arm picking a screwdriver off a bench and inserting it into a tool rack. Recorded on real hardware at 40 Hz with three RGB cameras, converted to LeRobot v3.0. This repository merges the original 36-episode collection with 30 additional DAgger-collected episodes (interactive corrections from a rolled-out policy's own visited states).

Dataset Details

Dataset Description

A green-handled screwdriver lies on a white bench; a black 12-hole tool rack sits at the back of the workspace. One arm grasps the screwdriver and inserts it into a hole in the rack. Both grippers actuate over essentially their full range, so the dataset exercises both hands rather than leaving one idle.

Each frame carries measured joint positions, the teleoperator's commanded joint positions, and three synchronised camera views. Every episode carries the same English instruction string.

Format	LeRobot v3.0
Episodes	66 (36 original + 30 DAgger)
Frames	189,312
Duration	78.9 min (41.1–127.1 s per episode, mean 71.7 s)
Control rate	40 Hz
Robot	YAM bimanual — 2 x 6-DoF arms + 1-DoF gripper each (14 DoF)
Cameras	3 x 480x640 RGB, AV1 / yuv420p, 40 fps
Size on disk	~2.7 GB

• Curated by: internal robotics team (see Dataset Card Contact)
• Funded by [optional]: [More Information Needed]
• Shared by [optional]: [More Information Needed]
• Language(s) (NLP): en — a single fixed instruction string; this is not a natural-language dataset
• License: MIT

Dataset Sources [optional]

• Repository: internal (yam-data-collection, branch yam/groot-lerobot); no public mirror
• Paper [optional]: none
• Demo [optional]: none

Uses

Direct Use

• Imitation learning / behaviour cloning for bimanual manipulation.
• Vision-language-action post-training. This data has been used to post-train GR00T N1.7 and Pi0.5.
• Benchmarking insertion-style tasks, where the terminal phase is spatially constrained and the grasp quality determines success.
• Studying how recovery/retry demonstrations affect policy robustness — roughly 5 of the original 36 episodes contain a deliberate missed grasp followed by a recovery. Additionally, the 30 DAgger episodes (episode indices 36–65) contain interactive corrections collected from states a rolled-out policy actually visited, rather than expert-only recoveries — see the note under Bias, Risks, and Limitations below.

Out-of-Scope Use

• Generalization to new object placements. The screwdriver starts in essentially the same spot every episode. A policy trained here has no basis for handling a moved object and should not be evaluated as if it does.
• Anything safety-critical without an independent safety layer. These are demonstrations, not a validated controller. Policies trained on this data have been observed commanding joint steps ~37% above the demonstrated ceiling when given out-of-distribution input.
• Natural-language research. There is exactly one instruction string. It carries no discriminative signal by design.
• Claims about multi-tool or cluttered-scene manipulation. One tool, one rack, one lighting condition, one operator.

Dataset Structure

Key	Shape	Dtype	Notes
observation.state	(14,)	float32	measured joint positions
action	(14,)	float32	commanded joint positions (teleop leader)
observation.images.egocentric	(3, 480, 640)	float32	view of the whole cell
observation.images.wrist_left	(3, 480, 640)	float32	left wrist camera
observation.images.wrist_right	(3, 480, 640)	float32	right wrist camera
timestamp, frame_index, episode_index, index, task_index	scalar		LeRobot standard

State and action share one 14-element ordering, with real joint names rather than generic indices:

left_yam_joint1 … left_yam_joint6, left_yam_gripper,
right_yam_joint1 … right_yam_joint6, right_yam_gripper

The names are load-bearing. Training code commonly selects joints by name — for instance excluding grippers from relative-action conversion by substring match. A dataset carrying generic joint_0..joint_13 trains to completion and yields a measurably worse policy without raising an error.

Arm joints are radians; grippers are normalised to [0, 1].

Provenance (DAgger episodes only). meta/provenance.parquet labels every frame of the 30 DAgger episodes (episode indices 36–65) with how it was produced:

Provenance	Frames	Meaning
demo	70,611	original teleoperated episodes (0–35)
robot	70,540	autonomous policy rollout
intv	29,443	human intervention / correction
preintv	11,875	window immediately preceding an intervention
home	6,843	return-to-home segments

Demonstrated ranges (merged, 66 episodes)

joint	min	max		joint	min	max
left_yam_joint1	−0.643	0.004		right_yam_joint1	−0.202	0.511
left_yam_joint2	0.004	2.402		right_yam_joint2	0.003	2.103
left_yam_joint3	0.006	2.710		right_yam_joint3	0.006	1.899
left_yam_joint4	−1.483	0.759		right_yam_joint4	−1.089	0.918
left_yam_joint5	−0.263	1.554		right_yam_joint5	−1.338	0.109
left_yam_joint6	−1.152	0.360		right_yam_joint6	−0.132	0.992
left_yam_gripper	0.075	0.999		right_yam_gripper	0.117	0.999

Peak per-step motion at 40 Hz, useful for setting deployment bounds: arm 0.0610 rad (p99.99 0.0373), gripper 0.1249 (p99.99 0.0874). (Measured on the original 36-episode collection; not yet recomputed against the merged 66-episode set.)

Splits. This repository is the training split. A 4-episode / 7,914-frame validation split was carved out of the second recording session and is not included here, so training here and evaluating there is a genuine holdout. The split is by whole episode, never mid-episode.

Dataset Creation

Curation Rationale

Collected to post-train and compare vision-language-action policies on a contact-rich insertion task using a small amount of real robot data — testing how far a few tens of minutes of demonstrations can be pushed, rather than assembling a large corpus.

The second session was recorded specifically because a policy trained on the first 15 episodes missed the grasp on 2 of 5 real-robot attempts and never retried: no episode in that set contained a retry, because the operator never missed while teleoperating. A behaviour absent from the data cannot be imitated.

A third round (episodes 36–65) followed the DAgger (Dataset Aggregation) protocol: a trained policy was rolled out on the real robot, and the operator intervened to correct it whenever it drifted or was about to fail. Unlike the expert-only retries in the original 36 episodes, these corrections originate from states the policy itself visits — directly addressing the "retries are expert recoveries from expert-visited states" limitation noted below.

Source Data

Data Collection and Processing

Bilateral teleoperation: an operator moves leader arms, follower arms track, and both streams are logged. Robot state is sampled at ~240 Hz and cameras at 30 Hz; both are resampled to a common 40 Hz row rate at conversion. Video is encoded AV1 / yuv420p.

Cameras are addressed by USB serial rather than /dev/videoN, because device numbers are reassigned on replug and recording against a stale number silently fills a view with the wrong camera.

Recorded across five sessions on the same physical setup:

Session	Episodes contributed	Notes
session_20260820_212015	15	clean first-try picks only
session_20260821_185006	21 (of 25; 4 held out)	includes deliberate retries
session_20260824_203132	3	DAgger round: policy rollout + intervention
session_20260824_210417	7	DAgger round: policy rollout + intervention
session_20260824_222406	20	DAgger round: policy rollout + intervention

Episodes aborted during recording are excluded at conversion. Retry episodes widen the demonstrated workspace considerably — left_yam_joint3 reaches +2.735 rad in the original 36-episode set against +2.008 in the first session alone, since recovering from a missed grasp sweeps far more of the workspace than a clean pick.

Who are the source data producers?

A single human teleoperator, operating the leader arms directly (both during teleoperated demonstration and during DAgger interventions). No demographic information is collected or included.

Annotations [optional]

Annotation process

No post-hoc annotation. The only label is a single constant instruction string applied to every episode at conversion time:

put the screwdrivers into the tool rack

The plural is deliberate — the collection plan adds more tools later, and a constant string across sessions means a policy conditions on vision rather than on language.

Who are the annotators?

Not applicable; no annotation pass was performed.

Personal and Sensitive Information

No personal data is deliberately included: no faces, names, or identifying metadata, and the operator's hands are outside the camera views for the most part.

However — the egocentric camera captures the surrounding room. Before any public release, review that stream for incidental content (whiteboards, screens, documents, people passing) and confirm the recording location may be shown publicly.

Bias, Risks, and Limitations

1. Near-fixed object placement. The screwdriver starts in essentially the same spot every episode. No positional diversity.
2. Single scene, single lighting condition, single operator. No domain randomization of any kind.
3. Original-collection retry behaviour is sparse — roughly 5 of the first 36 episodes. Enough to widen the demonstrated workspace, likely not enough on its own to teach reliable recovery.
4. The DAgger round (episodes 36–65) partially addresses this: those 30 episodes are interactive corrections from states a rolled-out policy actually visited, rather than expert-only recoveries from expert-visited states. This dataset card has not yet been updated with a measured before/after generalization or recovery-rate comparison — treat the DAgger addition as a more targeted signal, not a proven fix.
5. ~79 minutes total is still small. Sufficient to produce a working policy on a fixed scene (a GR00T N1.7 policy trained on the 15-episode subset grasped on 3 of 5 real attempts) but not for generalization.
6. Measured overfitting ceiling (on the original 36-episode set). On a GR00T N1.7 fine-tune, held-out arm MAE stayed flat at 0.0110–0.0117 rad across all ten checkpoints while training-set arm MAE fell 0.0056 → 0.0024 — the gap widening from 2.0x to 4.5x. About 4,000 optimizer steps bought the same generalization as 20,000. Not yet re-measured against the merged 66-episode set.
7. Video is AV1. Some decoders lack hardware AV1 support; torchcodec under a fork multiprocessing context has been seen to fail with Could not push packet to decoder: Invalid data found when processing input.

Recommendations

• Hold episodes out and rank checkpoints on them. Without a holdout only the training curve is visible, it falls monotonically, and the final checkpoint always appears best — an artifact of memorization, not evidence of skill.
• Do not rank on open-loop error alone. On the original 36-episode set, held-out error was flat across all checkpoints while held-out behaviour still differed sharply: one checkpoint left the demonstrated workspace on 72.2% of ticks against 43.2% for another with identical error.
• Bound deployment by the demonstrated workspace, not the mechanical range, using the ranges above, and give arms and grippers separate step ceilings — they are in different units, and throttling the gripper costs grasps.
• Declare 40 fps at conversion, not the 30 fps camera rate. Declaring 30 stretches every timestamp by a third.
• Expect this dataset to be a floor, not a ceiling. The DAgger round adds on-policy correction data as recommended previously; further rounds, more placement diversity, or more interventions remain the likeliest path to better robustness.

Citation [optional]

BibTeX:

@misc{patel2026yamtoolrack,
  author = {Patel, Kevin},
  title  = {YAM bimanual --- tool-rack placement},
  year   = {2026},
  note   = {Teleoperated and DAgger-corrected demonstrations, LeRobot v3.0 format}
}

APA:

Patel, K. (2026). YAM bimanual — tool-rack placement [Data set]. Teleoperated and DAgger-corrected demonstrations, LeRobot v3.0 format.

Glossary [optional]

• LeRobot v3.0 — Hugging Face's robot-dataset format: Parquet tables for low-dimensional streams, separate encoded video per camera, metadata in meta/.
• Episode — one complete attempt at the task, from home pose to completion.
• Action vs state — action is what the teleoperator commanded; state is what the arms measured. They differ by tracking error.
• Relative actions — training-time conversion of absolute joint targets into per-step deltas. Grippers are normally excluded and kept absolute.
• Holdout / held-out split — episodes excluded from training, used to measure generalization rather than fit.
• Open-loop error / MAE — teacher-forced prediction error against recorded actions. Measures fit; does not measure task success.
• DAgger (Dataset Aggregation) — a data collection protocol where a trained policy is rolled out and a human corrects it live; the resulting corrections are added back to the training set so the policy sees data from its own visited states, not just expert demonstrations.
• Provenance — per-frame label (demo, robot, intv, preintv, home) recording how a DAgger-round frame was produced; see Dataset Structure above.

More Information [optional]

Companion write-ups exist for the two post-training runs performed on this data (GR00T N1.7 and Pi0.5), covering throughput, checkpoint selection, and on-robot results. They are internal and not distributed with this dataset.

Dataset Card Authors [optional]

[More Information Needed]

Dataset Card Contact

[More Information Needed]