AutoNav Showcase (static, read-only)

A · Research overview

Can LiDAR odometry keep up with a 10 Hz sensor?

A reproducible study measuring whether scan-to-scan LiDAR odometry can meet a 10 Hz sensor's 100 ms deadline, what it costs in accuracy, and what it takes to close the gap, built before committing to robot hardware.

It is a measurement study, not a navigation system. Every result is labelled with the hardware it ran on.

Sensor period
100 ms

1 / 10 Hz. A frame that takes longer delays every later frame.

Primary dataset
OS1-64

Newer College 2020, Ouster OS1-64 Gen1, 64 beams, 10 Hz, with ground truth.

Baseline median latency
750 ms

Development laptop, no limits. Measured on 13th Gen Intel(R) Core(TM) i5-13420H (laptop-full); not a Raspberry Pi, not embedded hardware.

Current answer

Verdicts below are generated by documented rules from the run's measurements (20261110T080635Z_newer_college_bag9_baseline_laptop-full_0862f4f). They describe this machine and this dataset only.

Did the pipeline meet the 100 ms deadline?

FAIL

No. 100.0 % of 798 frames missed 100 ms. Median latency 750 ms is 7.5x the deadline; P99 is 2112 ms.

Rule and evidence

pass: 0 violations; warn: 0 < violations <= 5 %; fail: violations > 5 %. Violation = per-frame pipeline latency (preprocess + register) strictly greater than the deadline.

{
  "deadline_ms": 100,
  "p50_ms": 749.8913,
  "p95_ms": 1547.597945,
  "p99_ms": 2111.641545999998,
  "max_ms": 2684.1567,
  "violation_percent": 100,
  "frames_measured": 798
}
Did registration produce valid matches?

PASS

269 of 269 registrations passed the validity gate (100.0 %).

Rule and evidence

pass: >= 99 % accepted; warn: >= 90 %; fail: < 90 %. Accepted = passes the registration validity gate.

{
  "attempts": 269,
  "accepted": 269,
  "rejected": 0,
  "rejected_by_reason": {},
  "icp_iterations_median": 10,
  "inlier_ratio_median": 0.99936,
  "residual_median_m": 0.12463
}
How accurate was the estimated trajectory?

INFO

ATE RMSE 0.51 m (max 1.06 m) over 22.5 m of ground-truth path after rigid alignment; position RPE is reported per window in the summary. Both include any unmodelled sensor-to-body offset (lever arm) between the estimate and the ground-truth body frame, so they mix drift with extrinsic bias. Scan-to-scan odometry has no loop closure, so error accumulates with distance.

Rule and evidence

descriptive; no pass/fail threshold is applied to accuracy

{
  "ate_rmse_m": 0.5068686215458409,
  "ate_max_m": 1.0565121346857196,
  "gt_path_length_m": 22.472984119767467,
  "ate_rmse_percent_of_path": 2.2554575700518296,
  "final_error_m": 0.19201671534112752,
  "rpe_mode": "world_position_after_umeyama_alignment",
  "rpe": {
    "time_1s": 28.56793327591613,
    "time_5s": 14.567642505063228,
    "distance_5m": 10.553605423564546,
    "distance_10m": 7.834970613283673
  }
}

How the question is answered

Pipeline

DatasetPreprocessICPGatePoseEvaluate Python writes run artifacts (JSON/CSV/binary). This site only renders them.

Each stage is timed separately, in a fresh process per repeat, after warm-up frames are excluded.

Method

  • Scan-to-scan point-to-point ICP with a validity gate; accuracy against ground truth where it exists.
  • Latency percentiles and the deadline violation rate, not just a percentile.
  • Three algorithm variants side by side: baseline, optimized, point-to-plane.
  • Hardware profiles: full laptop, 4 cores, 1 core. ARM cloud and Raspberry Pi 5 are listed separately.

Scope and limits