Comparing Trajectories from Positions Alone: Curvature-Based Time Alignment and Drift Error Metric

Impact Assessment

Core Contribution. This paper addresses a genuine and under-served problem in field robotics: how to fairly evaluate estimated trajectories when the only available "ground truth" is a stream of 3D positions (from GNSS or a Robotic Total Station), without reliable orientation, time synchronization, or extrinsic calibration. The standard metrics — ATE and RPE — either require full 6-DoF reference poses (RPE) or absorb error through global alignment (ATE), and both are highly sensitive to under-reported evaluation parameters. The paper contributes three linked elements: (i) a curvature-signature-based temporal alignment method that recovers the clock offset between estimated and reference trajectories via cross-correlation of discrete curvature signals; (ii) a Drift Error (DE) metric defined as the windowed ratio of travelled distances, requiring only positions and correctable for lever-arm (extrinsic) errors via the ratio of curvature radii; and (iii) an integrated evaluation protocol validated with case studies on the GrandTour and FoMo datasets.

Methodological Rigor. The theoretical grounding is sound and elegant: the observation that all points on a rigid body share the same instantaneous center of rotation, so curvature radius varies synchronously and is invariant to rigid transforms, is a clean justification for orientation-invariant temporal alignment. The lever-arm correction via radius ratios is derived carefully with a proper treatment of degenerate cases. However, the empirical validation is deliberately framed as "case studies, instead of large-scale evaluation." This is honest but limits the strength of the evidence. The time-calibration demonstration rests essentially on one mission (GrandTour HEAP-1), the lever-arm demonstration on a manually injected 0.5 m offset over 5 m of one trajectory, and the sensitivity analysis on nine trajectories from each of two datasets. There is no statistical comparison across many algorithms, no error bars, and only two reference-noise levels to support the recommended ϵ_min = 3σ heuristic. Baselines are limited to the "temporal" (assume Δt=0) and Umeyama "spatial" alignment approaches, plus displacement error (KITTI-style point distance). This is adequate to illustrate the claims but not to establish them robustly.

Potential Impact. The practical value is real: trajectory benchmarking is ubiquitous in SLAM, odometry, and state estimation, and the position-only constraint captures the majority of large-scale outdoor/field datasets. A metric that works from positions alone and is robust to time and calibration errors could genuinely broaden the range of usable benchmark data and improve reproducibility. If adopted into a widely-used toolkit (e.g., evo), the influence could be substantial. That said, the DE metric is conceptually close to the existing KITTI relative-drift-over-segment percentage and the "displacement/point distance" metric it compares against; the incremental novelty is the curvature-based lever-arm correction and the temporal alignment, rather than the drift concept itself. The paper also candidly notes that both contributions require curvature excitation and degrade on straight-line motion — a meaningful limitation for highway or corridor scenarios.

Timeliness & Relevance. The topic is timely: the community (Zhang & Scaramuzza, Lee & Civera) has recently converged on the view that ATE/RPE are insufficient and that evaluation parameters are under-reported. This paper fits that active conversation and extends it to the position-only field-robotics regime that prior alternative metrics (DTE/DRE, mAA, TAS/RAS/PAS) do not handle because they presuppose orientation.

Strengths. (1) A clear, physically-motivated insight (ICR/curvature invariance) applied to two problems at once. (2) Explicit, honest treatment of assumptions — synchronization, sampling, extrinsic calibration — that are usually swept under the rug. (3) Practical, actionable parameter-selection guidance. (4) Validation on real, recent datasets rather than simulation, which the authors argue (fairly) is more convincing for a metrics paper.

Limitations. (1) Evidence is illustrative rather than comprehensive; claims like "65% error inflation from a 70 ms offset" rest on single examples. (2) No released code, which is ironic for a paper advocating reproducibility and standardization — adoption depends heavily on a public, well-maintained implementation. (3) The straight-line degradation is a genuine restriction. (4) DE discards orientation error entirely, so it cannot fully replace RPE/ATE where orientation matters; it is complementary. (5) Novelty relative to KITTI-style drift metrics is moderate.

Other observations. The work is low-cost and easily extensible by any robotics lab using public datasets, which favors uptake. The theoretical framing is the paper's strongest asset and gives it a chance to become a referenced building block if the protocol and an implementation are disseminated. The paper partially corroborates Zhang & Scaramuzza's finding that reported error depends on the evaluation protocol, and mildly challenges over-reliance on ATE/RPE, but does not overturn any load-bearing belief.

Overall, this is a solid, well-motivated methods paper solving a real pain point with an elegant geometric idea, but its impact is capped by limited empirical validation, moderate conceptual novelty relative to existing drift metrics, absence of released code, and inherent restriction to curvature-rich trajectories.