Local-first, deterministic validation for physical-AI recordings and robot-learning datasets.
Quickstart · Demo · Rules · Evidence · Roadmap · Contributing
Physlint finds concrete integrity defects before robot data reaches training. It explains the impact, identifies the affected episode and stream, recommends remediation, writes a stable JSON report, and returns a CI-safe exit code.
The 0.1.0a1
public alpha ships with a publicly validated LeRobot Dataset v3.x adapter. The engine is designed to grow into recording and dataset profiles for MCAP/ROS 2, Robomimic HDF5, RLDS/TFDS, and other physical-AI formats without pretending those adapters already exist.
Important
Physlint validates configured data-integrity contracts. A pass does not certify policy quality, task success, or robot safety.
Catch failures before GPU time: malformed manifests, broken episode ranges, reordered clocks, missing values, corrupt video, frozen cameras, and black frames become actionable findings.Keep robot data local:physlint check
performs no network requests and never modifies its source dataset.Get evidence, not a mystery score: every finding includes a stable rule ID, severity, source location, observed condition, expected condition, impact, and remediation.Use it in CI: deterministic execution, versioned JSON, atomic report writes, strict configuration, and documented exit codes.Trust scoped claims: the public release gate pins exact dataset revisions and commits sanitized reports, corruption recipes, checksums, and publication metrics.
Physlint requires Python 3.11 or newer.
python -m pip install "physlint[video]==0.1.0a1"
To test the exact tagged source instead of the PyPI distribution, install the GitHub release directly:
python -m pip install "physlint[video] @ git+https://github.com/pranayjoshi/physlint.git@v0.1.0a1"
physlint inspect /path/to/lerobot-dataset
physlint check /path/to/lerobot-dataset
Write JSON to an exact destination:
physlint check /path/to/lerobot-dataset \
--output json \
--json-output artifacts/physlint-report.json
The source remains untouched. Exit code 0
means the configured contract passed; 1
means validation completed with a blocking finding.
These captures use the pinned Panda source from the release gate. The second dataset is a fully dereferenced copy with one deterministic NaN injected at episode 0, sample 5, state dimension 0.
| Clean pinned snapshot | Controlled NaN corruption |
|---|
Physlint owns NaN and infinity semantics in numeric.finite_values
; the same sample is not duplicated as a missing-stream finding.
Seventeen deterministic rules are enabled by default:
| Area | Checks |
|---|---|
| Manifest | |
| Required files, declared/stored schema agreement, required streams, and feature shapes | |
| Episodes | |
| Unique identifiers, positive lengths, non-overlapping boundaries, and stored-row agreement | |
| Temporal | |
| Strictly monotonic timestamps, FPS cadence, FPS-aware maximum gaps, complete stream overlap, and observation/action delay when independently timestamped | |
| Numeric | |
| NaN/Inf, configured physical bounds, and configured discontinuity limits | |
| Video | |
| Complete decode, motion-aware frozen-frame runs, and grouped black/near-empty frames |
List or explain the installed rule contract:
physlint rules
physlint rules --json
physlint explain temporal.max_gap
physlint explain video.frozen_frames
Rules whose required inputs are unavailable return not_run
with a reason; they are never misreported as passed. Robot-specific bounds and discontinuity checks stay not_run
until the user supplies meaningful thresholds.
Read the complete MVP rule specifications.
Run physlint init
to generate a documented quality contract, or create physlint.yaml
yourself:
config_version: 1
adapter: auto
required_streams:
- observation.state
- action
fail_on: error
rules:
temporal.max_gap:
options:
max_gap_multiplier: 2.0
video.frozen_frames:
options:
max_consecutive_frames: 5
motion_streams: [action, observation.state]
numeric.configured_bounds:
options:
limits:
action:
min: [-1.0, -1.0]
max: [1.0, 1.0]
numeric.discontinuity:
options:
max_delta:
observation.state: [0.25, 0.25]
reports:
json: true
output_dir: .physlint/reports
Use it explicitly when needed:
physlint check /path/to/dataset --config physlint.yaml
Unknown top-level keys, rule IDs, rule options, duplicate required streams, and invalid values are rejected instead of silently ignored.
The CLI has stable exit codes and writes reports atomically, so a basic GitHub Actions gate is small:
- name: Install Physlint
run: python -m pip install "physlint[video]==0.1.0a1"
- name: Validate robot dataset
run: |
physlint check "$DATASET_PATH" \
--json-output artifacts/physlint-report.json
- uses: actions/upload-artifact@v4
if: always()
with:
name: physlint-report
path: artifacts/physlint-report.json
| Exit code | Meaning |
|---|---|
0 |
|
| Validation completed and the configured contract passed | |
1 |
|
| Validation completed and the contract failed | |
2 |
|
| Invalid command or configuration | |
3 |
|
| Dataset or adapter failure | |
4 |
|
| Internal Physlint error | |
130 |
|
| Interrupted by the user |
The alpha release gate evaluates four immutable public snapshots from four producers:
| Dataset | Robot | Episodes | Frames | Applicable rules | Result |
|---|---|---|---|---|---|
ViaCatalyst/robomimic-can-ph-lerobot-v3 |
cagataydev/scout-earth-rover-mini-20260616-053232
lerobot/svla_so101_pickplace
vikram-avea/sentinel-demo-09
Clean-source result: 4/4 snapshots pass with zero findings and zero rule errors. Controlled-defect recall: 3/3 for a non-finite value, reordered timestamps, and a deleted source row.
Everything needed to audit or rerun those claims is versioned:
Pinned repository manifestDeterministic corruption and execution harnessSanitized reports and SHA-256 valuesPublication-ready CSVManual classification and performance analysisReproduction instructions
Runtime measurements are observations from the documented machine and run—not universal performance guarantees.
The storage format is an adapter boundary, not the product boundary.
| Format | Status | Intended mode |
|---|---|---|
| LeRobot Dataset v3.x | Alpha—implemented and publicly validated | |
| Training datasets | ||
| MCAP with ROS 2 profiles | Planned—seeking design partners | Recordings and derived datasets |
| Robomimic HDF5 | Planned | Demonstration datasets |
| RLDS/TFDS | Researching | Episode/step datasets |
| ROS bag2 SQLite and ROS 1 bag | Researching | Recordings |
MCAP needs two honest layers: container/channel health that works without training semantics, and an explicit profile mapping topics to actions, state, cameras, and episode boundaries. See the cross-format roadmap and MCAP/ROS proposal.
Use the adapter-request issue form to contribute an immutable public example and a real failure mode.
Supported:
- LeRobot v3.x
meta/info.json
schema and path templates - Chunked Parquet episode metadata and sample shards
- Multiple episodes per shared Parquet/MP4 file
- Fixed-size and regular vector features
- Shared video segments using per-camera timestamp ranges
- Metadata-first discovery and bounded batch iteration
Not currently supported:
-
LeRobot v2.0/v2.1
-
Remote Hub identifiers passed directly to
physlint check -
Image-directory features in the video rule set
-
Arbitrary codecs unavailable to the installed OpenCV build
-
Inferred safety, calibration, task-success, or coordinate-frame conclusions
Read the LeRobot adapter boundary.
source format → read-only adapter → canonical episodes/streams → capability planner
↓
deterministic rule engine
↓
terminal + versioned JSON evidence
Read only: source datasets are never repaired or rewritten.Lazy by default: metadata first, bounded Parquet batches, and one shared privacy-safe video analysis pass.Explicit applicability: adapters advertise capabilities; unavailable checks explain why they did not run.Stable evidence: rule versions, fingerprints, source revisions, and report schema are serialized.Exception isolation: one rule failure cannot masquerade as a clean dataset pass.
Validation is offline. Reports contain source references, timestamps, aggregate statistics, and targeted evidence—not embedded images or complete source samples. Treat every dataset parser as an attack surface and report suspected vulnerabilities privately through GitHub Security Advisories.
See SECURITY.md before submitting a vulnerability. Do not attach private datasets or sensitive reports to public issues.
Contributions are welcome, particularly:
- Public healthy and defective datasets for adapter release gates
- False-positive reproductions
- MCAP/ROS recording schemas and failure modes
- New deterministic rules with controlled corruptions
- Documentation, performance characterization, and privacy reviews
Development setup:
git clone https://github.com/pranayjoshi/physlint.git
cd physlint
python -m pip install -e ".[video,dev]"
ruff check .
ruff format --check .
mypy
pytest
Rules require positive and negative fixtures, stable remediation, a bounded finding count, and controlled corruption evidence where applicable. Adapters must remain read-only, metadata-first, lazy over samples, and explicit about capabilities.
Read CONTRIBUTING.md, open a format request, or join GitHub Discussions.
Physlint is an alpha. Its current claims are deliberately limited to the documented LeRobot v3 boundary and committed release evidence. The project does not train policies, repair data, host datasets, infer task success, produce an opaque quality score, or certify that a robot or policy is safe.
See CHANGELOG.md for release notes.
Physlint is available under the MIT License.