Scalable Behavior Cloning with Open Data, Training, and Evaluation
Paper • 2606.27375 • Published
adapter dict | checks list | dataset dict | episodeCount int64 | findings list | frameCount int64 | inputCount int64 | limitations list | modalities list | provenance dict | quality dict | readiness dict | recommendations list | schemaVersion string | source dict | status string | targetFormatVersion string | taskCount int64 | warnings list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
{
"actionTopic": "/left-arm-action.plot",
"episodeCount": 1,
"frameCount": 136,
"source": "episode.fo.mcap",
"sourceSynchronization": {
"deltaMsMax": 0,
"deltaMsP50": 0,
"deltaMsP95": 0,
"pairedMessages": 136,
"pairingMethod": "message_index",
"unpairedActionMessages": 0,
"unpaired... | [
{
"category": "metadata",
"evidence": {
"episodes": 1,
"fps": 20,
"frames": 136,
"framesRead": null,
"requiredColumns": null,
"episodesChecked": null,
"longestEpisodeFrames": null,
"shortestEpisodeFrames": null,
"featuresChecked": null,
"framesChec... | {
"episodeCount": 1,
"format": "lerobot",
"formatVersion": "v3.0",
"fps": 20,
"frameCount": 136,
"inputCount": 1,
"name": "abc130k-screwdriver-mcap",
"robotType": "ABC-130k bimanual teleoperation",
"taskCount": 1
} | 1 | [
{
"category": "training_quality",
"confidence": "medium",
"evidence": {
"outlierFrameRatePct": 37.5
},
"id": "observation-state-outliers",
"message": "Some frames are more than six median absolute deviations from the feature median.",
"recommendation": "Inspect the affected signal ... | 136 | 1 | [
"Automated readiness checks do not measure downstream policy performance or task success.",
"Action saturation uses observed signal behavior unless hardware command limits are supplied separately.",
"Visual blur, exposure, and frozen-frame findings are heuristics and should be confirmed with representative epis... | [
{
"codec": null,
"depthUnit": null,
"dtype": "float32",
"feature": "action",
"fps": null,
"kind": "action",
"pixelFormat": null,
"shape": [
12
]
},
{
"codec": null,
"depthUnit": null,
"dtype": "float32",
"feature": "observation.state",
"fps": null,
... | {
"adapter": {
"revision": "4230bc4b56abe85116b202398b1a625a81c1c55c",
"type": "mcap",
"version": "1.0.0"
},
"sourceDataset": "Voxel51/ABC-130k",
"sourceLicense": "apache-2.0",
"sourceRevision": "9659e8ce4b39580f48369cc31bc2e47a217c40e7"
} | {
"dataset": {
"durationSeconds": 6.8,
"episodeCount": 1,
"fps": 20,
"frameCount": 136
},
"episodes": {
"durationSeconds": {
"max": 6.8,
"mean": 6.8,
"min": 6.8,
"p50": 6.8,
"p95": 6.8
},
"lengthFrames": {
"max": 136,
"mean": 136,
"min": ... | {
"findingCounts": {
"critical": 0,
"high": 0,
"low": 0,
"medium": 2
},
"label": "Review recommended",
"score": 84,
"scoreMethod": "Severity-weighted findings; raw warnings are reported but not double-counted. This is not a model-performance guarantee.",
"summary": "The dataset is structural... | [
"Inspect the affected signal for unit changes, resets, clipping, or corrupt samples."
] | 2.0 | {
"actionKey": null,
"actionTopic": "/left-arm-action.plot",
"adapterType": "mcap",
"adapterVersion": "1.0.0",
"media": [],
"mode": "adapter_discovery",
"source": "episode.fo.mcap",
"sourceSynchronization": {
"deltaMsMax": 0,
"deltaMsP50": 0,
"deltaMsP95": 0,
"pairedMessages": 136,
"... | passed | v3.0 | 1 | [] |
ABC-130k MCAP — verified LeRobot v3 reference episode
Community conversion produced by ViaCatalyst BYOD. This repository is not an official upstream release and is not affiliated with ABC-130k or Voxel51.
This is a provenance-complete conversion of one pinned MCAP episode from Voxel51's Apache-2.0 repackaging of ABC-130k. It is a reproducible reference for MCAP numeric-topic selection and LeRobot v3 compatibility.
At a glance
| Property | Value |
|---|---|
| Input format | MCAP |
| LeRobot format | v3.0 |
| Robot | ABC-130k bimanual teleoperation |
| Task | Put the screwdriver in the bin |
| Episodes / frames | 1 / 136 |
| Duration / FPS | 6.8 seconds / 20 |
| State / action dimensions | 26 / 12 |
| Selected state topic | /left-arm-state.plot |
| Selected action topic | /left-arm-action.plot |
| Official LeRobot reader | Passed with LeRobot 0.6.0 |
Feature mapping and action preservation
| LeRobot feature | Source |
|---|---|
observation.state |
Numeric fields from /left-arm-state.plot: end-effector pose, joint position, torque, and velocity |
action |
Numeric fields from /left-arm-action.plot: end-effector pose and joint position |
| Episode and frame indices | Single source-log episode and message ordering |
timestamp |
Output frame_index / 20 seconds; source pairing evidence is retained in validation-report.json |
Action messages are copied directly from the selected source topic. The conversion does not replay a policy, regenerate commands, interpolate actions, or relabel the episode. The source MCAP includes camera data, but the current MCAP adapter intentionally converts only the documented numeric telemetry topics; no visual modality is claimed by this release.
Intended use
- Validating MCAP numeric-topic selection and LeRobot v3 readers
- Reproducible testing of ROS-adjacent telemetry conversion workflows
- Small offline imitation-learning or educational experiments after reviewing validation evidence
Validation evidence
All seven critical checks passed, including schema/counts, timeline, finite feature values, LeRobot v3 metadata, provenance, and the official LeRobot reader smoke test. The readiness score is 84/100 — Review recommended. The report retains two medium-confidence robust-outlier signals (37.5% for state and action) for user review; they are not proof of corruption.
Read the complete evidence in validation-report.json.
Source, integrity, and license
- Source:
Voxel51/ABC-130k, an Apache-2.0 modified subset of ABC-130k - Pinned revision:
9659e8ce4b39580f48369cc31bc2e47a217c40e7 - Source file:
data/val/put_the_screwdriver_in_the_bin/episode_02eb2789-4234-4d4c-984e-08872ed00c1d/episode.fo.mcap - Source SHA-256:
f605f24fc0c3edee1549dd47a8934b48d68a700ee61fc93093e7815e0b058069 - Adapter:
mcap1.0.0; converter revision4230bc4b56abe85116b202398b1a625a81c1c55c
Audit artifacts: provenance.json, bundle-manifest.json, bundle-manifest.external.json, and UPSTREAM_LICENSE.md.
The source card declares Apache-2.0. This conversion retains the license, upstream attribution, and citation; it does not transfer ownership or replace source terms.
Limitations
- This is one MCAP episode, not the full ABC-130k corpus.
- It contains numeric telemetry only; it is not the source's full camera-plus-telemetry multimodal representation.
- State/action streams are paired by message index; review synchronization evidence before control training.
- Validation does not measure downstream policy quality or task success.
Load with LeRobot
from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset("ViaCatalyst/abc130k-screwdriver-mcap-lerobot-v3")
print(dataset.meta.total_episodes, dataset.meta.total_frames)
Conversion tooling
Converted and validated with the ViaCatalyst BYOD Processing Platform, a free workflow for converting robotics data to LeRobot format. For high-volume datasets, contact ViaCatalyst support through the platform.
Citation
@misc{abc2026,
title={Scalable Behavior Cloning with Open Data, Training, and Evaluation},
author={Allshire, Arthur and Singh, Himanshu Gaurav and Singh, Ritvik and others},
year={2026},
eprint={2606.27375},
archivePrefix={arXiv},
primaryClass={cs.RO},
doi={10.48550/arXiv.2606.27375}
}
- Downloads last month
- 14