README.md

metadata

license: cc-by-nc-4.0
task_categories:
  - robotics
language:
  - en
size_categories:
  - 100K<n<1M
tags:
  - robotics
  - teleoperation
  - imitation-learning
  - manipulation
  - lerobot
  - so-101
  - vla
  - world-model
  - laboratory-automation
  - test-tube
  - nesting
  - GR00T
  - Isaac-compatible
  - tabular
  - timeseries
  - video
  - parquet
  - dask
pretty_name: DecisionFacts Physical AI — SO-101 Test Tube Nesting
configs:
  - config_name: default
    data_files: data/*/*.parquet

DecisionFacts Physical AI — SO-101 Test Tube Nesting

Real-world, human-teleoperated demonstrations of an SO-101 arm nesting test tubes, recorded in LeRobotDataset v3.0 format with two synchronized RGB cameras, joint state and action at 30 fps. This repository is a public evaluation sample of the DecisionFacts Physical AI catalog.

Robot	SO-101 follower (so_follower), 5-DoF arm + gripper, single embodiment
Episodes	121
Frames	179,127 (≈ 100 minutes at 30 fps)
Episode length	≈ 54 s on average, 60 s maximum
Tasks	2 language-annotated task variants (see below)
Cameras	cam_front, cam_top — 640×480 RGB, 30 fps, AV1
Format	LeRobotDataset v3.0 (Parquet + MP4)
Split	train only (episodes 0–120)
License	CC BY-NC 4.0 (research and evaluation)

Task

The operator picks up test tubes and nests them into <TARGET — e.g. a rack / a larger tube> on a tabletop workspace. The task stresses skills that matter for lab-automation and fine-manipulation policies:

• Precise grasping of thin, cylindrical objects with a small gripper footprint.
• Tight-tolerance insertion: the tube must be aligned with the opening before release, so small orientation errors cause failure.
• Multi-step sequencing within a single, roughly minute-long episode, which makes the data useful for long-horizon policy and world-model evaluation.

Language instructions (from meta/tasks.parquet):

task_index	Instruction	Share of frames
0	<TASK 0 INSTRUCTION — copy verbatim from meta/tasks.parquet>	≈ 40%
1	<TASK 1 INSTRUCTION — copy verbatim from meta/tasks.parquet>	≈ 60%

Collection protocol

• Expert teleoperation. Each episode is a goal-directed demonstration by a trained operator using a leader arm; no scripted or simulated motion.
• Synchronized capture. Joint state, commanded action and both camera streams are recorded in lock-step at 30 fps.
• Review before release. Episodes are checked for completeness, timing alignment and coherent execution of the intended task before release.

Data structure

meta/
  info.json            # schema, fps, codec, paths
  stats.json           # per-feature normalization statistics
  tasks.parquet        # task_index -> language instruction
  episodes/chunk-000/  # per-episode metadata (length, file offsets, stats)
data/chunk-000/        # frame-level Parquet (state, action, indices)
videos/
  observation.images.cam_front/chunk-000/*.mp4
  observation.images.cam_top/chunk-000/*.mp4

Features

Feature	Type	Shape	Description
observation.state	float32	6	Measured follower joint positions
action	float32	6	Commanded joint positions from the leader arm
observation.images.cam_front	video	480×640×3	Front view of the workspace
observation.images.cam_top	video	480×640×3	Overhead view of the workspace
timestamp	float32	1	Seconds since the start of the episode
frame_index	int64	1	Frame number within the episode
episode_index	int64	1	Episode ID (0–103)
index	int64	1	Global frame ID
task_index	int64	1	Key into meta/tasks.parquet

Joint order for observation.state and action:

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

Arm joints are in degrees; the gripper is on a 0–100 scale (0 = closed). Observed ranges (from meta/stats.json):

Joint	State min	State max	State mean
shoulder_pan	−87.6	47.6	−2.9
shoulder_lift	−105.1	74.5	−7.5
elbow_flex	−81.1	99.3	30.3
wrist_flex	−96.1	99.5	7.8
wrist_roll	−131.5	165.9	77.9
gripper	1.7	57.1	15.8

The commanded gripper value is typically lower than the measured one while an object is held (median 7.9 vs 13.1), which is expected: the leader commands a tighter grip than the follower can reach around a tube.

Video encoding

Both streams are AV1 (yuv420p, CRF 30, GOP 2), decoded with PyAV or TorchCodec. The short GOP keeps random frame access fast during training. If your pipeline cannot decode AV1, re-encode to H.264 with ffmpeg -c:v libx264.

Usage

LeRobot

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("DecisionFacts/Physical_AI_SO101_Test_Tube_Nesting_Task")
print(ds.num_episodes, ds.num_frames, ds.fps)

frame = ds[0]
frame["observation.state"]              # tensor(6)
frame["action"]                         # tensor(6)
frame["observation.images.cam_front"]   # tensor(3, 480, 640)
frame["task"]                           # language instruction

Train a policy (example: ACT):

lerobot-train \
  --dataset.repo_id=DecisionFacts/Physical_AI_SO101_Test_Tube_Nesting_Task \
  --policy.type=act \
  --output_dir=outputs/test_tube_act

Tabular data only (no video)

from datasets import load_dataset

ds = load_dataset("DecisionFacts/Physical_AI_SO101_Test_Tube_Nesting_Task", split="train")

Visualize

Browse episodes in the LeRobot dataset visualizer.

Intended uses

• Imitation-learning baselines (ACT, Diffusion Policy) for precise pick-and-insert.
• Fine-tuning Vision-Language-Action models (e.g. SmolVLA, π0) on a laboratory-style manipulation task.
• Evaluating world models and video-prediction models on long-horizon, contact-rich manipulation.

Limitations

• Single embodiment and setup. All data comes from one SO-101 arm in one workspace with fixed camera placement; expect a drop in performance under new lighting, backgrounds or camera poses.
• Two camera views, RGB only. No depth or wrist camera; transparent tubes can be hard to localize.
• Joint-space only. No end-effector poses, forces or tactile signals.
• Evaluation sample. 104 episodes are enough for fine-tuning and benchmarking, not for training a general policy from scratch.
• No predefined validation split. Hold out episodes yourself (for example the last 10%).

License and access

This repository is released under CC BY-NC 4.0 for research and evaluation. Commercial use, redistribution, or training models intended for commercial deployment requires a separate license (see below).

Licensing the full catalog

This repository is a sample of a larger, continuously growing catalog. Related public samples on SO-101:

• Cup Nesting
• Towel Folding
• Stationary Sorting
• Pipe and Joint Sorting

We offer:

• Pilot tranches (200–500 episodes) for evaluation and fine-tuning experiments
• Production tranches (1,000+ episodes) for training and post-training at scale
• Custom collection for new tasks, environments, diversity requirements or other embodiments

Contact info@decisionfacts.io to scope access.

Citation

@misc{decisionfacts_so101_test_tube_nesting_2026,
  title        = {DecisionFacts Physical AI: SO-101 Test Tube Nesting},
  author       = {Raghav, Prabhu and Unni, Sreeram B and Gopalan, Sriram and Kumar, Saravana},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/DecisionFacts/Physical_AI_SO101_Test_Tube_Nesting_Task}}
}