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

arXiv:2609.14936 · cs.RO · Submitted 2026-09-14 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Comparing Trajectories from Positions Alone".

Dev: Here's the gist of the paper. In field robotics, acquiring independent largescale reference trajectories more accurate than the evaluated estimates remains an open challenge.

Rosa: This paper introduces a trajectory-evaluation protocol for standardized and reliable accuracy assessment in state estimation, localization, and Simultaneous Localization And Mapping (SLAM).

Dev: The approach combines a novel temporal alignment method based on curvature signals with an error metric normalized by travelled distance.

Rosa: The Problem with Existing Metrics Traditional metrics like Absolute Trajectory Error (ATE) and Relative Pose Error (RPE) depend on evaluation parameters that are rarely documented consistently, making comparisons across studies potentially unreliable.

Paper summary: Rosa: So thinking about this paper Comparing Trajectories from Positions Alone: Curvature-Based Time Alignment and Drift Error Metric, we’re looking at what this actually means for how we evaluate robotics systems in the real world. The authors are trying to move away from relying on metrics that depend on assumptions about setup parameters.

Dev: They introduce a protocol that combines temporal alignment based on curvature signals with the Drift Error metric, which is an error expressed with respect to the distance traveled. This metric can be corrected for the effect of a lever arm using a relation derived from local curvature radius, where alpha equals reference radius over estimated radius.

Rosa: It’s about mitigating two specific problems: getting the timing synchronized and having a distance-normalized error that isn't overly sensitive to missing calibration information. The way they handle temporal misalignment can change the relative ranking of algorithms in benchmarks if you don't correct for it first.

Taro: If you ignore that time offset, the performance comparison between two different localization methods can be completely skewed by how well those methods happen to align temporally with your noisy reference data.

Dev: The authors show that using the Drift Error metric helps reduce the impact of a missing extrinsic calibration information, for example they added a manual zero point lateral offset and saw a thirty-seven percent error for state of the art displacement error compared to forty-point five percent for their metric on three dimensional positions.

Rosa: That difference shows how much better this new way of measuring distance error is at being less sensitive to that kind of calibration gap when you're only using three dof positions. It's a concrete improvement over what was previously available.

Taro: The limitation they mention is that both the curvature excitation and the Drift Error metric itself degrade when you are looking at straight-line segments, which means these methods might not be as effective in purely linear motion scenarios.

Dev: So it’s a tool for more rigorous comparisons where you have to account for time and calibration uncertainties explicitly, even if it means adding a little complexity to the evaluation pipeline.

Rosa: For the listener driving or walking, this means that when you look at data from different sources, you can now have more confidence in which system is actually better because the comparison itself is standardized and less dependent on perfect prior knowledge of every setup detail.

Taro: It moves us toward a future where we can trust trajectory evaluation even when the field conditions are messy, as long as we apply this protocol to account for those inherent real-world uncertainties.

Conclusion: Rosa: So this paper is all about how to compare robot paths when you don't have perfect timing or perfect reference data.

Dev: Yeah, they introduce this new protocol called Drift Error that tries to fix those issues by looking at the path distance itself instead of just relying on absolute positions.

Rosa: It seems like they’re tackling the problem of getting independent long-distance reference paths that are actually usable outside a controlled lab setting.

Taro: What they propose is using curvature signals to figure out the time offset between two trajectories because rigid bodies have this property where every point shares a center of rotation at one specific time.

Dev: They estimate that time lag by finding the best match for the curvature signatures, which is basically a way to align them based on how fast they are turning.

Rosa: And they define this Drift Error as just the difference between the distance traveled by your estimated path and what you expect from a reference path, without needing to line up every single point perfectly in space.

Taro: They even have this correction factor using local curvature radius, which basically adjusts the error based on how much the turning radius of your estimate differs from your reference.

Dev: And they show that this metric is pretty robust to issues like a missing lever arm, meaning if you don't know exactly where the reference system was positioned relative to your robot, it doesn't completely break the error calculation.

Rosa: That’s interesting because it means we can start evaluating systems on datasets that aren't perfectly curated or perfectly calibrated yet.

Taro: But they do mention a limitation, which is that both using the curvature signals and this Drift Error metric actually gets worse when you are looking at straight lines.

Dev: So, while it’s great for curves and turns, if your robot is just driving in a straight line for a long time, these methods might not give you as much of an advantage.

Rosa: It suggests that the choice of evaluation metric really depends on what kind of motion you're actually testing.

Taro: Exactly. So, knowing when to use this protocol versus sticking to older metrics is going to be a big part of how we test autonomy in the real world.

Effie Daum, Daniele De Martini, Claire Dune, Francois Pomerleau

Northern Robotics Laboratory, University of Laval · Mobile Robotics Group, Oxford Robotics Institute

cs.RO

Submitted: 2026-09-14

Updated: 2026-11-03

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 78/100

The gist: In field robotics, acquiring independent largescale reference trajectories more accurate than the evaluated estimates remains an open challenge.

Key concepts

Curvature-Based Temporal Alignment
This method estimates the time difference between two paths by analyzing their curvature signals. It uses the idea that points on a rigid body must share a common center of rotation at any given time. By finding the lag that maximizes the correlation of these curvature signatures, it recovers the true temporal offset.
Drift Error (DE) Metric
The DE metric quantifies estimation error based purely on distance traveled, ignoring spatial alignment issues. It is calculated as 1 minus the ratio of estimated distance to reference distance. A correction factor based on local curvature radius helps account for how different radii affect the actual path length during rotation.
Lever Arm Correction
This correction accounts for errors introduced by missing extrinsic calibration, such as a lever arm in rotational motion. The metric adjusts the error calculation using a relation derived from local curvature radius to mitigate the impact of this missing information on the final error estimate.

Terminology

Summary

In field robotics, acquiring independent largescale reference trajectories more accurate than the evaluated estimates remains an open challenge. p. 0

Where the paper says it

Comparing Trajectories from Positions Alone: Curvature-Based Time Alignment and Drift Error Metric Effie Daum 1 , Daniele De Martini 2 , Claire Dune 1 , Franc¸ois Pomerleau 1 Abstract— In field robotics, acquiring independent largescale reference trajectories more accurate than the evaluated estimates remains an open challenge. The domain is widely reliant on Absolute Trajectory Error (ATE) and Relative Pose Error (RPE), computed with automated tools, that rest on assumptions and evaluation parameters rarely made explicit.

This paper introduces a trajectory-evaluation protocol for standardized and reliable accuracy assessment in state estimation, localization, and Simultaneous Localization And Mapping (SLAM). p. 0
Where the paper says it

When unreported, the errors can be misleading and hinder fair comparisons. This paper introduces a trajectory-evaluation protocol for standardized and reliable accuracy assessment in state estimation, localization, and Simultaneous Localization And Mapping (SLAM).

The approach combines a novel temporal alignment method based on curvature signals with an error metric normalized by travelled distance. p. 0
Where the paper says it

This paper introduces a trajectory-evaluation protocol for standardized and reliable accuracy assessment in state estimation, localization, and Simultaneous Localization And Mapping (SLAM). The approach combines a novel temporal alignment method based on curvature signals with an error metric normalized by travelled distance.

The Problem with Existing Metrics

Traditional metrics like Absolute Trajectory Error (ATE) and Relative Pose Error (RPE) depend on evaluation parameters that are rarely documented consistently, making comparisons across studies potentially unreliable. ATE depends on the global alignment between estimated and reference trajectories, which may absorb part of the estimation error. p. 0

Where the paper says it

However, their reported values depend on evaluation parameters that are rarely documented consistently, making comparisons across studies potentially unreliable [6]. ATE depends on the global alignment between estimated and reference trajectories, which may absorb part of the estimation error.

RPE requires full 6-Degrees-of-Freedom (DoF) reference poses. p. 0
Where the paper says it

ATE depends on the global alignment between estimated and reference trajectories, which may absorb part of the estimation error. RPE requires full 6-Degrees-of-Freedom (DoF) reference poses [7], whereas large-scale outdoor reference systems such as Global Navigation Satellite System (GNSS) or Robotic Total Station (RTS) typically provide only 3D positions, as illustrated in Fig.

Furthermore, large-scale outdoor reference systems typically provide only 3D positions, necessitating reconstruction of missing orientations which compromises groundtruth independence. p. 0
Where the paper says it

RPE requires full 6-Degrees-of-Freedom (DoF) reference poses [7], whereas large-scale outdoor reference systems such as Global Navigation Satellite System (GNSS) or Robotic Total Station (RTS) typically provide only 3D positions, as illustrated in Fig. Reconstructing the missing orientations using additional sensors, such as an Inertial Measurement Unit (IMU) [8], compromises groundtruth independence when the same sensors contribute to the evaluated estimator [2].

Curvature-Based Temporal Alignment

The authors propose a curvature-based temporal alignment method to recover the time offset between trajectories directly from their timestamps and positions. This method relies on the fundamental property of rigid body motion that enforces that all points on that rigid body must share the same Instant Center of Rotation (ICR) at a given time. p. 2

Where the paper says it

Therefore, we propose a solution to recover ∆t directly from the estimated and reference trajectories and their timestamps. Our solution relies on the fundamental property of rigid body motion that enforces that all points on that rigid body must share the same Instant Center of Rotation (ICR) at a given time.

The curvature is estimated using the formula κi = 1/ri = 2∆θ/∆s1 + ∆s2, where ∆θ is the angular change subtended over approximated arc length ∆s1 and ∆s2. p. 2
Where the paper says it

Moreover, it is invariant to rigid transformations, making the time calibration procedure robust to wrong spatial calibration. For each trajectory sample i, the curvature κi is estimated over an ϵmin-neighbourhood: κi = 1 ri = 2∆θ ∆s1 + ∆s2 , (1) where ∆θ is the angular change subtended over approximated arc length ∆s1 and ∆s2 as illustrated in Fig.

The temporal offset is then estimated as the lag that maximizes the normalized discrete crosscorrelation of curvature signatures ρ(k). p. 3
Where the paper says it

κ ref, respectively, we define their normalized discrete crosscorrelation for a candidate lag k as ρ(k) = X i∈Ik κ est i κ ref i+k sX i∈Ik κ est i 2 X i∈Ik κ ref i+k 2 , (2) where Ik denotes the set of indices for which both curvature signatures are defined. The temporal offset between the two trajectories is then estimated as the lag that maximizes the correlation k ⋆ = arg max k∈[−kmax,kmax] ρ(k), ∆t = k ⋆ δt, (3) where δt is the sampling period.

The Drift Error Metric (DE)

The proposed Drift Error (DE) metric quantifies the estimation error with respect to the distance travelled, comparing path lengths without any spatial alignment needed. DE is defined as the ratio between the travelled distances, DE = 1 - ∆s est / ∆s ref. p. 3

Where the paper says it

, GNSS or RTS) provide the reference trajectory. Given two time-aligned trajectories, DE is defined as the ratio between the travelled distances, DE = 1 − ∆s est ∆s ref , (4) where ∆s est and ∆s ref denote the path lengths travelled by the estimated and reference trajectories, respectively, over the same window length w ∈ R + in meters.

The metric can be corrected for the effect of the lever arm using a relation derived from local curvature radius, where α = r ref / r est. p. 3
Where the paper says it

Their theoretically travelled distances ∆s ∗ therefore differ only through their respective radii, yielding the ratio α: α = ∆s ref* ∆s est* = r ref∆θ r est∆θ = r ref r est . (5) Using this relation, DE can be corrected for the effect of the lever arm through the following more general formulation: DE = 1 − α ∆s est ∆s ref .

This correction is based on the arc-length relation and accounts for how different radii result in different travelled distances during rotational motion. p. 3
Where the paper says it

, ATE, RPE) as well as DE, since different lever arm magnitudes result in different travelled distances during rotational motion. In our framework, the local curvature radius provides a direct means to correct this effect using the arc-length relation.

Validation and Sensitivity Analysis

The protocol was validated on GrandTour [6] and FoMo [9] datasets. Case studies demonstrated the impact of time calibration, showing that temporal misalignment can alter the relative ranking of algorithms. p. 3

Where the paper says it

Therefore, we propose different case studies, instead of large-scale evaluation, to guide researchers needing to evaluate multiple algorithms. Case study: Impact of time calibration To illustrate the significant impact of even a small temporal misalignment and evaluate our alignment method, we compare two trajectories from the HEAP-1 mission of the publicly available GrandTour dataset [6] with regard to a reference trajectory.

Furthermore, the DE metric proved robust to errors caused by a lever arm, with its impact being less severe on DE than on standard displacement error. p. 4
Where the paper says it

Case study: impact of lever arm on DE evaluation metrics We want to demonstrate the capability of DE to be robust to errors caused by a lever arm (i. , a error in the extrinsic calibration of the sensors).

Sensitivity evaluations showed that the minimum displacement parameter ϵmin acts as a low-pass filter for noise in the reference trajectory. p. 5
Where the paper says it

First, we evaluate the impact of the minimum displacement ϵmin parameter. This parameter is used in the computation of the curvature and acts as a low-pass filter to counteract the noise in the reference trajectory.

The window length w and overlap β do not directly impact the error in DE but change how it is reported. p. 6
Where the paper says it

Finally, unlike the previous parameter studies, the window length w and the window overlap β used to split a trajectory do not directly impact the error in the process and are not specific to DE. The window length w changes how the error is reported.

Conclusion

The proposed protocol using Drift Error can mitigate the impact of wrong time and extrinsic calibrations, therefore opening evaluations on less curated datasets. The authors showed that if the time offset is left uncorrected, the ranking in the evaluation process can be altered and impact algorithm performance in benchmarks. p. 7

Where the paper says it

process can be altered and impact algorithm performance in benchmarks. Our proposed evaluation metric Drift Error (DE) can reduce the lever arm impact from missing extrinsic calibration information, where we added a manual 0.

The proposed evaluation metric Drift Error (DE) can reduce the lever arm impact from missing extrinsic calibration information, where they added a manual 0. p. 7
Where the paper says it

process can be altered and impact algorithm performance in benchmarks. Our proposed evaluation metric Drift Error (DE) can reduce the lever arm impact from missing extrinsic calibration information, where we added a manual 0.

5 m lateral offset, reporting 37 % for the state-of-the-art displacement error for 3-DoF positions compared to 4. p. 7
Where the paper says it

Our proposed evaluation metric Drift Error (DE) can reduce the lever arm impact from missing extrinsic calibration information, where we added a manual 0. 5 m lateral offset, reporting 37 % for the state-of-the-art displacement error for 3-DoF positions compared to 4.

5 % for our metric. p. 0
Where the paper says it

Comparing Trajectories from Positions Alone: Curvature-Based Time Alignment and Drift Error Metric Effie Daum 1 , Daniele De Martini 2 , Claire Dune 1 , Franc¸ois Pomerleau 1 Abstract— In field robotics, acquiring independent largescale reference trajectories more accurate than the evaluated estimates remains an open challenge. The domain is widely reliant on Absolute Trajectory Error (ATE) and Relative Pose Error (RPE), computed with automated tools, that rest on assumptions and evaluation parameters rarely made explicit.

Both contributions require curvature excitation and degrade on straight-line segments. p. 7
Where the paper says it

We analyzed the sensitivity of the parameters for the curvature computation and DE metric using the GrandTour dataset [6] and the FoMo dataset [9], standardizing the evaluation protocol. Both contributions require curvature excitation and degrade on straight-line segments.

The authors propose that a reasonable guess for ϵmin is three times the expected standard deviation of a reference trajectory. p. 5
Where the paper says it

05 m and 0. Therefore, even if we only have two noise levels to corroborate our observations, we propose that a reasonable guess for ϵmin is three times the expected standard deviation of a reference trajectory.

The authors suggest setting the window length w while keeping enough meaningful observation points. p. 6
Where the paper says it

For a high window length w, we are confident that we observe the drift of the estimated reference, but we reduce the number of observations heavily, thus being more sensitive to outliers. We suggest setting the window length w while keeping enough meaningful observation points.

The authors suggest setting a = 0. p. 6
Where the paper says it

For a high window length w, we are confident that we observe the drift of the estimated reference, but we reduce the number of observations heavily, thus being more sensitive to outliers. We suggest setting the window length w while keeping enough meaningful observation points.

1 meter for both datasets. p. 6
Where the paper says it

For this reason, we proposed a = 0. 1 meter for both datasets.

ACKNOWLEDGMENT

Generative AI (Claude, Anthropic) was used to assist in generating the plotting code for the data-driven result figures in Section IV; the underlying data, analysis, and figure content were produced and verified by the authors.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Temporal alignment improves trajectory comparison by using a curvature-based method that estimates time offset from trajectory curvature signatures, which is less sensitive to drift and less demanding on speed estimation compared to methods relying on crosscorrelation of speed profiles or matching nearest timestamps.

  2. The Drift Error (DE) metric allows for robust evaluation of trajectories from 3D positions alone by comparing the ratio between the travelled distances without requiring full 6-DoF poses, addressing a limitation where standard metrics like RPE are undefined.

  3. The DE metric corrects for lever arm impact by using the arc-length relation to find a correction factor alpha, leading to a more accurate comparison of trajectories even when extrinsic calibration is unknown or inaccurate.

  4. The evaluation protocol provides explicit accounting for temporal synchronization and extrinsic calibration through a sensitivity analysis, allowing researchers to quantify their influence on reported error values and enabling more rigorous, reproducible, and standardized trajectory evaluation.

Abstract

In field robotics, acquiring independent large-scale reference trajectories more accurate than the evaluated estimates remains an open challenge. The domain is widely reliant on Absolute Trajectory Error (ATE) and Relative Pose Error (RPE), computed with automated tools, that rest on assumptions and evaluation parameters rarely made explicit. When unreported, the errors can be misleading and hinder fair comparisons. This paper introduces a trajectory-evaluation protocol for standardized and reliable accuracy assessment in state estimation, localization, and Simultaneous Localization And Mapping (SLAM). The approach combines a novel temporal alignment method based on curvature signals with an error metric normalized by travelled distance. We explicitly account for temporal synchronization, sampling alignment, and extrinsic calibration, quantifying their influence through a sensitivity analysis. The proposed protocol contributes to more rigorous, reproducible, and standardized trajectory evaluation.

Sources

Related papers