Start with two comparable episodes
RunFP compares stored robot episodes to show changes in captured evidence, including available environment and behavior signals. Use a baseline and candidate from the same task with compatible recording and evaluation conditions.
Create the episodes using the getting started guide, then replace these placeholders with actual episode IDs:
runfp diff <baseline-episode-id> <candidate-episode-id>
The comparison report helps inspect what changed across recorded runs. Depending on the signals present, behavior evidence can include joint states, odometry, poses, and actuator values.
Set the reference for repeated checks
runfp baseline set <baseline-episode-id> --name pick-place
runfp history pick-place
The named baseline is reused by group checks. History shows attempts with available commit and outcome evidence, which can narrow an investigation to relevant software changes.
Apply your task's criteria
Use a prepared protocol to make required evidence and tolerances explicit:
runfp diff <baseline-episode-id> <candidate-episode-id> --protocol protocols/pick-place.yaml
Review missing topics, incompatible structures, and protocol findings before interpreting metric differences. A recording without the signal you need cannot establish that behavior stayed the same.
For an automated comparison gate, choose a finding threshold deliberately:
runfp diff <baseline-episode-id> <candidate-episode-id> --fail-on high
Thresholds refer to the comparison engine's findings. They are not a universal definition of a successful physical task. CI checks and recorded task outcomes answer different release questions.
What the evidence establishes
| Evidence | What you can use it for | What still needs validation |
|---|---|---|
| Captured software or model changes | Identify differences worth investigating | Whether a change caused a failure |
| Recorded motion or state differences | Inspect observed behavior changes | Whether behavior meets the physical task objective |
| Missing or incompatible signals | Identify limits on comparability | Recovering evidence that was never recorded |
| An associated commit | Trace a run to available workspace provenance | Reproduction when dependencies or local edits differ |
A difference report supports investigation; it does not automatically establish root cause. If the recorded evidence points to a field failure, request an OORB reliability audit.