How should a robot budget latency around an edge-vision model?

Measure capture, transfer, preprocessing, inference, postprocessing and delivery separately, then measure the complete path. Preserve the source capture timestamp instead of replacing it with the time inference finishes. If a platform moves at 0.5 m/s, a 200 ms-old observation corresponds to 100 mm of travel. That age matters even when the model itself is fast.

Choose the queue policy according to the task. For live obstacle perception, processing every old frame can create a growing backlog; dropping stale frames may be preferable if the downstream system can handle the reduced rate. For offline inspection, preserving every frame may be required. Track queue depth and observation age so a low inference time is not mistaken for low end-to-end latency.

Keep coordinate frames and timestamps consistent when publishing detections to robot software. Test behavior when the camera stalls or inference fails, and ensure the motion controller has an appropriate response to stale perception. Benchmark under the expected CPU, GPU and network load. A single unloaded timing measurement does not characterize the worst relevant operating condition.