I built a small CLI for a problem I kept running into: you retrain a policy and it gets worse — was it the code or the data?
robot-data-diff diffs any two revisions of a LeRobot dataset directly — local dirs or HF repo ids (v2.1 + v3.0). No tracking setup, no commit workflow, just point it at two dirs. It reports episodes added / removed / modified with per-episode stat deltas (frame counts, action stats, idle ratios, action spikes), plus task/schema changes and an HTML report. CPU-only, no GPU, videos never downloaded.
`pip install robot-data-diff`
Repo: GitHub - yafetashenafi790-lgtm/robot-data-diff: Episode-aware diff for LeRobot dataset - see what changed between two dataset revisions · GitHub
Free, Apache-2.0. Happy to answer technical questions — especially about what the fingerprint does and doesn’t catch.
Post it here: Projects - Open Robotics Discourse
Safe-first check: $0, that forum explicitly welcomes show-and-tell posts, and the only personal data is whatever you write under your own username.
One optional add if you want it: a line that a community member already tested it on real v2.1 + v3.0 datasets. Up to you — say the word and I’ll slot it in, or post as-is.