Reproduce robot failures with rosbag2: inputs, TF and time

Audit a fixed replay in your browser

Use the panel above to compare clock settings, timestamp offsets, reliability QoS and TF selection. The Lab computes a synthetic replay audit, not ROS, DDS, tf2, a localizer or a robot simulation. It opens no bag files and connects to no robot. The ROS commands below are optional follow-up procedures.

Every run uses eleven scan records at bag times 10.0–12.0 s, spaced 0.2 s apart. Recorded input odom → base_link samples span 9.8–12.2 s at the same spacing. These are invented teaching fixtures, not hardware measurements. The timestamp control adds an offset to scan headers only; bag recording times and TF stamps stay fixed. It represents a differently stamped sensor fixture, not a rosbag2 playback option. Elapsed playback time is (bag time − 10 s) / rate. At 2×, the scan span becomes 1 s while the headers remain in their original time domain.

Read the four checks separately

1. Modeled receipt: choose the publisher's offered and subscriber's requested reliability independently. Best-effort offered to reliable requested is incompatible; the other three combinations are compatible. Topic names/types and every other QoS policy are assumed compatible. All compatible scans arrive in this loss-free model. This does not promise real best-effort delivery, establish actual rosbag2 auto-selected QoS, or model queues, deadlines, lifespan, discovery or late subscribers.
2. Recorded TF coverage: the scan header must fall within the recorded input TF interval, including endpoints. This is an audit of the complete recording, even when the modeled subscriber receives no scans. It does not simulate tf2's live cache, arrival ordering, interpolation, waiting, or transform availability at a callback. A stamp inside the interval is therefore not proof that a live lookup succeeds.
3. Clock difference: the replay-clock fixture sets the node's latest /clock value equal to each scan's bag time, assuming a clock update arrives before each scan. The system-clock fixture starts at the arbitrary value 1000 s and advances by elapsed playback time. With use_sim_time but no received clock, node time is zero and the diagnostic is not evaluated, shown as a dash. For initialized clocks, the Lab counts |node now − scan header| ≤ 150 ms. That symmetric window is a teaching diagnostic, not a ROS validity rule, tf2 tolerance or localizer timeout. It is independent of receipt and TF coverage.
4. TF publishers: the evaluated localizer is configured to publish map → odom. “Input transforms only” keeps recorded odom → base_link; “all recorded transforms” adds a second map → odom publisher; “exclude all recorded TF” removes the required input edge. Counts describe configured sources, not a computed localization result. Duplicate-source behavior and static-TF durability are not simulated.

Save the default as A and try B with 2× playback, a +400 ms header offset, no clock, a reliability mismatch, or a different TF selection. The plots use common A/B axes. With +400 ms, the last scan is beyond the recorded TF interval; all eleven scans lie outside the teaching clock window. With a missing clock, that window's count is unavailable rather than zero. Neither finding predicts trajectory error. The data table and JSON retain each scan's recording time, header, elapsed time, node time and independent check outcomes.

Reference and scope

The model's rules were reviewed on 2026-09-20 against the rosbag2 Jazzy README, Jazzy player options, Jazzy QoS documentation source, ROS clock design, Jazzy tf2 time tutorial source and REP-105 frame roles. No ROS installation was executed for this Lab.

Each A/B run starts fresh. Scheduling, packet loss, estimator memory, backward time jumps, multiple clock publishers, calibration, static TF, map accuracy and closed-loop driving remain outside this exercise. Use the SLAM evaluation Lab for a separate trajectory-error exercise. Browser checks do not replace the real replay checks below.