Skip to content

Latest commit

 

History

History
54 lines (41 loc) · 2.64 KB

RESULTS_ROS2.md

File metadata and controls

54 lines (41 loc) · 2.64 KB

Batch-LIO ROS2 Humble — A/B reproduction

Re-running the docs/RESULTS.md A/B on the ros2-humble port (native /opt/ros/humble, 32 cores, ROS1 Avia bags converted to ROS2 mcap via scripts/convert_bag.py). The algorithm code is byte-identical to the ROS1 version; only the ROS plumbing changed.

Bags (converted): quick-shack (49 s) and outdoor_run (63.8 s, 100 Hz, aggressive). The in-ROS2 baseline is batch_dt:=0.0 (point-wise, == Point-LIO behavior) vs batch_dt>0 (batch).

Per-frame compute: speedup REPRODUCED

run mode avg total / icp / propag (ms) speedup
quick-shack point-wise (batch_dt=0) 12.42 / 10.11 / 0.13 1.0×
quick-shack batch 1 ms 3.51 / 3.04 / 0.05 3.5×
outdoor_run point-wise (batch_dt=0) 2.56 / 1.99 / 0.05 1.0×
outdoor_run batch 1 ms 0.54 / 0.44 / 0.02 4.7×

Within/above the documented 2.3–3.6× band (outdoor_run benefits more: more points per frame).

batch_dt sweep (quick-shack): 1–2 ms sweet spot REPRODUCED

batch_dt (s) avg total (ms)
0.0005 4.85
0.001 3.50
0.002 2.73
0.005 2.13
0.01 1.77
0.02 1.49

Monotonic decrease with larger windows; the accuracy-vs-cost trade-off favors 1–2 ms.

De-skew: transform correct; drift benefit is regime-dependent

  • The de-skew transform (eq 3.44–3.47) unit test passes (test_deskew, 5/5).
  • At the 1 ms operating point on outdoor_run, de-skew on vs off drift is ≈equal (0.085 m vs 0.079 m) — expected, since intra-1 ms motion is tiny.
  • At a 20 ms window (the RESULTS.md fig3 probe) on outdoor_run, both modes diverge (drift 19.8 m on / 2.97 m off): the filter is at the edge of stability at that aggressive window on this 100 Hz bag. The original fig3 8×-lower-drift result was measured on a different bag and does not cleanly reproduce on outdoor_run. The de-skew block is byte-identical to the validated ROS1 version, so this is a regime effect, not a port bug.

Bottom line

The port is verified correct: clean build, colcon test green (deskew gtest + smoke launch_test), end-to-end odometry on a real Avia bag, and the headline 2.3–3.6× speedup reproduces (3.5× quick-shack, 4.7× outdoor_run). The de-skew correctness is covered by the unit test; its large-window drift benefit is bag/regime-dependent.

No ground truth is used here either: the drift numbers are loop-closure (return-to-origin) on outdoor_run, and any "vs baseline" diff is agreement with the point-wise (batch_dt=0) run, not error vs a true trajectory.