YERP + rosbag2 snapshot catches ROS2 latency spikes
A new YERP layer triggers rosbag2 snapshots only when perception fails
ZC Liu (ZC502) has released YERP, a minimal “EvidenceFlow” layer that sits above rosbag2 in ROS2 perception stacks. While rosbag2 snapshot already keeps recent messages in memory, YERP decides when to dump them by detecting runtime pressure events such as p95/p99 latency spikes, output candidate surges, or confidence entropy jumps.
In a standalone YOLO edge pipeline test, YERP flagged a frame as evidence-worthy because post-processing dominated 28 % of the cycle time and produced 18 bounding boxes with entropy 2.51, even though the image itself looked ordinary. The resulting metadata.json plus the frame gives engineers reproducible data for debugging rare timing faults on AMRs without recording gigabytes of normal traffic.
- YERP (YOLO Edge Runtime Profiler) by ZC502 triggers rosbag2 snapshot only when perception latency spikes or box counts surge
- Records structured JSON evidence—preprocess_ms, inference_ms, postprocess_ms, box_count, confidence_entropy—so engineers can reproduce field failures
- Test case showed a normal-looking frame flagged because postprocess dominated 28 % of runtime and produced 18 boxes
Why It Matters
Cuts debugging from terabytes of logs to kilobytes of evidence, making rare robot perception faults reproducible and fixable in the field.