State Machines — Bimanual Warehouse Samples
11 teleoperated robot episodes collected in warehouse offices in Texas. The tasks include handling mailers, peeling adhesive backing strips, inserting an item into a mailer, and closing a cardboard box.
This is a sample of State Machines robotics data for prospective buyers to inspect before licensing a larger collection.
What's included
- 30 minutes 56 seconds of aligned data: 11 episodes, 92,817 robot-data samples at 50 Hz.
- Average episode length: 2 minutes 49 seconds.
- Three cameras: overhead, left wrist and right wrist; 640 × 480 RGB video at 30 fps.
- Two MakerMods Metal robot arms: 14 motor positions and 14 action targets per robot-data sample, including the two gripper motors.
- State: measured absolute motor positions in degrees.
- Action: absolute motor position targets actually sent to the robot, in degrees, held until the next command.
- LeRobot v3.0 format: MP4 videos, Parquet data, English task descriptions, and source timing and trim records.
Data notes
Video note: Some recordings contain a small black-and-white marker in the top-left corner. It was embedded during capture for latency diagnostics.
Robot states and sent action targets are resampled at 50 Hz. Camera videos are encoded at 30 fps, separately from the robot-data rate. Both use the host receipt clock. Each grid uses the latest available source sample, with a maximum age of 100 ms. Frames can repeat or be dropped; no motion interpolation is used.
LeRobot selects the nearest 30 fps video frame for each 50 Hz row. The encoded
frame timestamp differs by up to about 13.3 ms; this is not a guarantee of causal
camera selection or hardware-synchronized exposures. Original exposure latency
is uncalibrated. Source-frame mappings are included in provenance/.
Camera calibration is not included. Invalid or stale intervals are excluded.
One recording ended early; its valid segment is included.
Task descriptions are AI-generated drafts based on sampled frames. Task success has not been verified. No policy-training results are included.
The dataset passed LeRobot loading, video decoding and numeric-data checks.
Results are in meta/validation.json; joint ordering is in meta/info.json.
Source records and frame alignment are in provenance/. Audio and untrimmed
recordings are not included.
Load
Use a LeRobot version supporting v3.0. Authenticate with an account that has access to this private dataset.
from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset(
"statemachinesai/bimanual-tx-wh-samples",
tolerance_s=0.02, # Match 30 fps video to the 50 Hz robot timeline.
)
sample = dataset[0]
print(sample["task"], sample["observation.state"], sample["action"])
The explicit tolerance_s=0.02 is required for mixed-rate video loading.
For action chunks, use time offsets in multiples of 1/50 second.
The Hub may display 50 fps from the dataset timeline; the MP4 streams are 30 fps.
Contact and licensing
Contact State Machines for more data. Data collection for specific warehouse tasks is available on request.
This sample dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). You may use, adapt and redistribute it, including for commercial purposes, under the license terms. Credit State Machines, link to the license and indicate any changes. See LICENSE for the full terms.
This license applies to the sample dataset in this repository. Larger collections and custom data collection are available under separate agreements.
- Downloads last month
- 928