How vedevpatel/kinema works
Kinema audits LeRobot robot-learning datasets before training, loading local, zipped, or Hub data and returning check findings, a health score, and reports. The main workflow is user → CLI → dataset loading → registered quality checks → report rendering. Checks cover data integrity, motion and calibration, metadata and video, duplicates and outliers, and deployment readiness; RLDS reading and MCP access are additional capabilities. Edges reflect README statements or source-index references; for unsampled modules, connections are omitted or marked unverified rather than inferred as certain.
Main entry points: base.py Check framework · mcp_server.py MCP server · cli.py CLI · dataset.py LeRobot loader
Components 26
User entry
- CLI
src/kinema/cli.py - MCP server
src/kinema/mcp_server.py
Dataset loading
- LeRobot loader
src/kinema/dataset.py - RLDS reader
src/kinema/rlds.py
Quality checks
- Check framework
src/kinema/checks/base.py - Action checks
src/kinema/checks/actions.py - Calibration checks
src/kinema/checks/calibration.py - Idle and actuator
src/kinema/checks/idle.py - Duplicate checks
src/kinema/checks/duplicates.py - Metadata checks
src/kinema/checks/metadata.py - Timestamp checks
src/kinema/checks/timestamps.py - Stats drift
src/kinema/checks/stats_drift.py - Stats provenance
src/kinema/checks/stats_provenance.py - Outlier checks
src/kinema/checks/outliers.py - Visual quality
src/kinema/checks/visual.py - Deployment readiness
src/kinema/checks/deployment.py - RLDS checks
src/kinema/checks/rlds_checks.py - Stats helpers
src/kinema/stats.py
Results and integrations
- Dataset diagnosis
src/kinema/diagnose.py - Report model
src/kinema/report.py - Dataset probes
src/kinema/probes.py - Similarity search
src/kinema/similar.py - Audit history
src/kinema/history.py
Other components
- Dataset auditor
- Hugging Face Hub
- Google Cloud Storage
Connections 33
- Dataset auditor → CLI: runs audit
README.md - CLI → LeRobot loader: loads dataset
src/kinema/cli.py - CLI → Check framework: selects checks
src/kinema/cli.py - CLI → Report model: renders results
src/kinema/cli.py - CLI → MCP server: starts server
src/kinema/cli.py - LeRobot loader → Hugging Face Hub: fetches dataset
README.md - LeRobot loader → Stats helpers: uses helpers
src/kinema/dataset.py - Check framework → Action checks: registers checks
src/kinema/checks/base.py - Check framework → Calibration checks: registers checks
src/kinema/checks/base.py - Check framework → Idle and actuator: registers checks
src/kinema/checks/base.py - Check framework → Duplicate checks: registers checks
src/kinema/checks/base.py - Check framework → Metadata checks: registers checks
src/kinema/checks/base.py - Check framework → Timestamp checks: registers checks
src/kinema/checks/base.py - Check framework → Stats drift: registers checks
src/kinema/checks/base.py - Check framework → Stats provenance: registers checks
src/kinema/checks/base.py - Check framework → Outlier checks: registers checks
src/kinema/checks/base.py - Check framework → Visual quality: registers checks
src/kinema/checks/base.py - Check framework → Deployment readiness: registers checks
src/kinema/checks/base.py - Check framework → RLDS checks: registers checks
src/kinema/checks/base.py - Check framework → LeRobot loader: checks dataset
src/kinema/checks/base.py - Stats drift → Stats helpers: compares stats
src/kinema/checks/stats_drift.py - Calibration checks → Stats helpers: uses helpers
src/kinema/checks/calibration.py - Idle and actuator → Stats helpers: uses helpers
src/kinema/checks/idle.py - RLDS reader → Google Cloud Storage: reads shards
src/kinema/rlds.py - RLDS checks → RLDS reader: checks data
src/kinema/checks/rlds_checks.py - Dataset diagnosis → LeRobot loader: loads dataset
src/kinema/diagnose.py - Dataset diagnosis → Check framework: runs checks
src/kinema/diagnose.py - Dataset diagnosis → Report model: builds report
src/kinema/diagnose.py - MCP server → Dataset diagnosis: dispatches diagnosis
src/kinema/mcp_server.py - MCP server → Report model: uses reports
src/kinema/mcp_server.py - MCP server → Similarity search: uses similarity
src/kinema/mcp_server.py - MCP server → Dataset probes: uses probes
- MCP server → Audit history: uses history