The author introduces Trajlens, an open-source linter for the LeRobotDataset category on Hugging Face Hub, and reports results from auditing 100 random public datasets tagged with 'lerobot'. The audit revealed that only 19 datasets passed validation, while 13 failed due to specific upstream bugs and 47 encountered load errors or timeouts.
- Of the failing datasets, approximately 19% suffered from episode-frame corruption (v2.1 to v3.0) identified in LeRobot issue 2401.
- Roughly 3% of failures were caused by timestamp float drift, linked to LeRobot issue 3177.
- The tool is available via `pip install trajlens` and can lint a specific dataset in under a minute.
Running Trajlens allows maintainers to quickly identify if their data is affected by these known corruption issues, helping to ensure data quality within the open robot-learning ecosystem.