vial-sort-v1-static

Teleoperated demonstration dataset for a vial pick-and-place task on a Waveshare SO-101 6-DOF arm, recorded with LeRobot v0.5.2 (dataset format v3.0). Intended for training ACT imitation-learning policies.

---

Episode definition

One episode = one complete pick-and-place cycle:

1. Arm starts at home pose.
2. Gripper descends to the red vial in the left rack.
3. Vial is lifted clear, transported, and placed in position 3 of the right rack.
4. Arm retracts to home pose.

100 episodes total, collected across four 25-episode sessions with at least a 30-minute break between sessions.

Split	Episodes	Red vial start (left rack)
Session 1	0–24	Position 1
Session 2	25–49	Position 3
Session 3	50–74	Position 4
Session 4	75–99	Position 5

Mean episode duration: 26.5 s at 30 fps. Total frames: 65 404.

---

Camera setup

Three cameras, all running at 640 × 480 px, 30 fps:

Key	Type	Mount	Notes
observation.images.cam_top	Waveshare IMX335 USB	Top-down, above the workspace	Addressed by USB physical-port path, fourcc=MJPG
observation.images.cam_wrist	Waveshare IMX335 USB	Gripper-mounted	Addressed by USB physical-port path, fourcc=MJPG
observation.images.cam_side	Intel RealSense D4xx	45° side view, fixed to table	Addressed by serial 052622071016, warmup_s=3

Cameras and racks are taped at fixed positions for the entire campaign; lighting is locked. The visual distribution the model trains on is exactly the scene shown here.

---

Color × rack mapping

Object	Color	Rack	Position	Role
Target vial	Red	Left	1, 3, 4, or 5 (varies by session)	Grasped and moved
Distractor vial	Blue	Right	1 (fixed, never moves)	Visual distractor
Destination	—	Right	3 (fixed across all episodes)	Drop target

Positions 2 and 6 of the left rack are never used. Rack holes are discrete and widely spaced; ACT is not expected to interpolate to unseen positions.

---

Language annotation

Every episode carries the same fixed task string stored as dataset metadata:

Place the red vial in position 3 of the right rack.

This prompt is metadata only. The policy trained on this dataset is ACT, which conditions exclusively on camera images and joint state. It never reads the language annotation at training or inference time. The string is stored for compatibility with the LeRobot schema and for future language-conditioned policy experiments.

---

Joint state and action

6-DOF SO-101 follower arm (STS3215 motors). Columns in observation.state and action, in order:

shoulder_pan.pos  shoulder_lift.pos  elbow_flex.pos
wrist_flex.pos    wrist_roll.pos     gripper.pos

Units: degrees.

---

Known limitations

• Fixed lighting — recorded under locked artificial lighting; performance under different illumination is untested.
• Fixed arm base — the follower arm is bolted to the same table position for all episodes; no base-pose variation.
• Fixed camera extrinsics — all three cameras are taped down; any bump or rebuild shifts the visual distribution.
• Fixed rack positions — racks are taped to the table; absolute rack pose is part of the training distribution.
• Limited color set — only red (target) and blue (distractor) vials appear; the policy is not expected to generalise to other colors.
• Four discrete start positions — the left rack has 6 holes; only positions 1, 3, 4, and 5 appear in training data.
• Single destination — right-rack position 3 is always the drop target; the policy does not generalise over placement location.

---

Recommended use

Suitable for training and benchmarking ACT-style imitation-learning baselines on a tabletop pick-and-place task. Not intended for production deployment.

Evaluation should place the red vial in one of the four trained starting positions (left-rack 1, 3, 4, or 5). Positions 2 and 6 are out-of-distribution.

---

Validation

All 100 episodes passed automated data-quality checks (camera frame coverage, joint discontinuities, stream sync, language annotation). Per-session QC reports are included in the training repository under qc_reports/.

Validation script: qc_dataset_v3.py

python qc_dataset_v3.py \
    --dataset-root <path-to-dataset> \
    --out report.json --md report.md \
    --frame-drop-pct 60

---

License

CC-BY-4.0 — free to use with attribution.