Per-Trajectory Error Map in Robotics

Updated 3 July 2026

Per-Trajectory Error Map is a structured framework that represents uncertainty, error, or residuals along individual trajectories using probabilistic, set-based, and score-based methods.
It supports continuous, segmentwise, or stepwise evaluation, enabling diagnostics, adaptive planning, and safety certification in robotics, navigation, and control.
By unifying paradigms from likelihood-based SLAM to quantum error mitigation, it provides granular insights into error sources, estimator drift, and reliability over time.

A Per-Trajectory Error Map is a structured, temporally or spatially indexed representation of uncertainty, error, or residual—accompanied, in rigorous frameworks, by formal propagation of noise and disturbance models—constructed for each individual trajectory arising in tasks such as state estimation, path tracking, prediction, or system identification. This concept subsumes both absolute and relative notions of error; supports continuous-time, segmentwise, or stepwise evaluation; and can be assembled from probabilistic, set-based, or score-based methodologies. It underpins quantitative diagnostics, safety certification, adaptive planning, and benchmarking in robotics, navigation, control, and prediction.

1. Likelihood-Based Continuous-Time Error Mapping in SLAM

The probabilistic, continuous-time approach to per-trajectory error mapping in SLAM, as formulated by Anderson and Barfoot, defines the ground-truth trajectory as a piecewise Gaussian Process (GP) in the Lie algebra $\mathfrak{se}(3)$ , mapping $t \mapsto \barT(t)$ and yielding an associated covariance $\Sigma(t)$ at arbitrary time $t$ (Zhang et al., 2019). This construction generalized temporal association beyond nearest-neighbor snapping by employing GP regression conditioned on sparse ground-truth poses.

Two principal error metrics are recast as likelihoods:

Generalized relative error is constructed from pose-delta increments and evaluated via the negative log-likelihood:

$E_{\rm rel}(t_i) = \frac{1}{2} e_{r,i}^\top V_{r,i}^{-1} e_{r,i} + \text{const}$

where $e_{r,i} = \log((\barT_{i}^{-1}\barT_{i+1})^{-1}(T_{i}^{-1}T_{i+1}))$, and $V_{r,i}$ is the propagated increment covariance.

Generalized absolute error employs pose alignment and marginalization over the unknown alignment, which, up to first order in BCH, yields equivalence to the relative likelihood.

Sliding the query time $t$ throughout the trajectory, one generates a continuous error map $E(t)$ , with peaks identifying consistent estimator drift, sudden estimator failures, or temporal windows of reduced trustworthiness, directly informed by ground-truth uncertainty. This paradigm provides interpretability not just as “how much” error, but “when” and, by correlation with auxiliary data, “why” errors arise.

2. Localized Error Bounds and Per-Segment Error Mapping in Optimal Planning

In the context of discrete-continuous trajectory optimization, per-segment or per-edge error maps provide localized discretization error estimation aligned to the underlying physics (e.g., Zermelo navigation under a wind field) (Borndörfer et al., 2022). The total trajectory cost difference $\Delta T = T(\xi_G) - T(\xi_C)$ , where $t \mapsto \barT(t)$0 is the discrete solution and $t \mapsto \barT(t)$1 is the continuous optimum, admits both:

A priori (global) bounds, scaling as $t \mapsto \barT(t)$2 in the graph connectivity scale $t \mapsto \barT(t)$3.
Localized per-segment error expressions, via integral bounds involving local derivatives of the cost functional. For an edge $t \mapsto \barT(t)$4 and midpoint $t \mapsto \barT(t)$5, the local error term has the form:

$t \mapsto \barT(t)$6

where $t \mapsto \barT(t)$7 depend on $t \mapsto \barT(t)$8, $t \mapsto \barT(t)$9, $\Sigma(t)$ 0 at $\Sigma(t)$ 1, and $\Sigma(t)$ 2, $\Sigma(t)$ 3 are spatial and angular deviations from a local reference.

These $\Sigma(t)$ 4 are assembled into an error vector $\Sigma(t)$ 5, providing explicit, segment-level diagnosis of approximation quality and supporting adaptive refinement (e.g., increasing node density where $\Sigma(t)$ 6 exceeds a threshold) (Borndörfer et al., 2022).

3. Set-Based, Invariant-Based Per-Trajectory Error Certification

In robust trajectory tracking, a per-trajectory error map quantifies, via time-indexed set projections, the reachable error induced by bounded disturbances and model uncertainties (Schitz et al., 9 Mar 2026). The methodology centers on:

Modeling closed-loop error dynamics as a linear parameter-varying (LPV) system with both additive ( $\Sigma(t)$ 7) and state-dependent ( $\Sigma(t)$ 8) disturbances.
Computing an ellipsoidal robust positive invariant (RPI) set $\Sigma(t)$ 9 via Linear Matrix Inequalities (LMIs).
Projecting $t$ 0 to the position error subspace to obtain a time-varying, trajectory-following ellipsoid $t$ 1; for heading-dependent architectures, this ellipsoid rotates along the path via $t$ 2.
Using $t$ 3, one defines a continuous error tube $t$ 4 for online buffer certification.

The continuous per-trajectory error map $t$ 5 functions as a certified safety margin and can be integrated into motion planners and collision-avoidance systems, guaranteeing constraint satisfaction under all modeled uncertainties (Schitz et al., 9 Mar 2026).

4. Statistical and Scenario-Aware Temporal Error Mapping in Prediction

For deep-learning-based trajectory prediction, a per-trajectory error map with statistical guarantees is realized by temporal calibration of prediction intervals and reliability scores (Shu et al., 5 Dec 2025). The major steps are:

Project prediction and ground-truth into a reference-aligned Frenet frame.
Compute per-timestep nonconformity scores and calibrate marginal prediction intervals $t$ 6 via CopulaCPTS, enforcing $t$ 7-level coverage.
Apply a scenario-aware reliability discriminator, comparing interval half-widths to maximum-allowable thresholds $t$ 8, parameterized by mean absolute error and scenario class.
Assemble a temporal risk series $t$ 9, segmenting the trajectory into reliable/unreliable intervals by thresholding.

This per-trajectory temporal error map offers granular, scenario-adaptive uncertainty and reliability assessment for downstream planning modules, with explicit formal coverage via conformal calibration (Shu et al., 5 Dec 2025).

5. Combined Score-Based and Heatmap Error Mapping in Map Matching and Filtering

In hybrid map matching and tracking under map uncertainty, per-trajectory error maps can aggregate pointwise unlikelihoods, off-road probabilities, or spatial deviations, yielding both stepwise error sequences and spatial heatmaps (Murphy et al., 2018). The semi-interacting multiple model (sIMM) filter maintains joint HMM (on-road) and Kalman (off-road) state, with per-step error metrics including:

Negative log likelihood ratio of on-road vs. off-road modes,
Probability of being off-road,
Euclidean distance from best on-road and off-road hypotheses.

Aggregating per-step errors spatially via kernel smoothing yields $E_{\rm rel}(t_i) = \frac{1}{2} e_{r,i}^\top V_{r,i}^{-1} e_{r,i} + \text{const}$ 0, a map-level error heatmap that diagnoses persistent failure regions in the map (e.g., missing links or misassigned constraints), and, when averaged across trajectories, supports data-driven map refinement (Murphy et al., 2018).

Per-trajectory error maps in vision-based navigation are analytically constructed by propagating pose and motion error covariances along the trajectory, starting from first-principles modeling of sources including quantization, terrain geometry, DTM uncertainty, and camera motion (Kupervasser et al., 2011). The core steps are:

For each keyframe, assemble a per-step covariance $E_{\rm rel}(t_i) = \frac{1}{2} e_{r,i}^\top V_{r,i}^{-1} e_{r,i} + \text{const}$ 1 using the measurement and terrain uncertainty structure.
Propagate the 6D $E_{\rm rel}(t_i) = \frac{1}{2} e_{r,i}^\top V_{r,i}^{-1} e_{r,i} + \text{const}$ 2 error covariance with the transition and accumulation $E_{\rm rel}(t_i) = \frac{1}{2} e_{r,i}^\top V_{r,i}^{-1} e_{r,i} + \text{const}$ 3.
Extract temporally indexed RMS position and attitude errors $E_{\rm rel}(t_i) = \frac{1}{2} e_{r,i}^\top V_{r,i}^{-1} e_{r,i} + \text{const}$ 4.

These quantities are plotted vs. time to yield a detailed per-trajectory error map, which diagnoses regimes of accumulating or bounded error according to observable and unobservable modes (Kupervasser et al., 2011).

7. Trajectory-Indexed Error Maps in Quantum Error Mitigation

In quantum systems, a per-trajectory error map is realized through stochastic unravelings of the Lindblad equation, associating each quantum trajectory with an influence martingale $E_{\rm rel}(t_i) = \frac{1}{2} e_{r,i}^\top V_{r,i}^{-1} e_{r,i} + \text{const}$ 5, reflecting the (quasi-)probability required to invert the effect of known noise (Donvil et al., 2023). The corrected estimate for any observable is then computed as a weighted sum over all trajectories:

$E_{\rm rel}(t_i) = \frac{1}{2} e_{r,i}^\top V_{r,i}^{-1} e_{r,i} + \text{const}$ 6

thus assigning a per-measurement correction that, in aggregate, inverts the noise channel on a trajectory-by-trajectory basis, with sampling variance depending on the trajectory weights.

Per-Trajectory Error Maps therefore unify a spectrum of trajectory-aware error quantification paradigms—ranging from probabilistic negative log-likelihood fields, set-valued buffer-zone evolutions, localized error bounds, calibrated prediction intervals, filtering-based residuals, analytically propagated covariances, to quasi-probabilistic quantum weighting—each providing granular insight into error sources, reliability, and opportunities for model, data, or planning improvement (Zhang et al., 2019, Borndörfer et al., 2022, Murphy et al., 2018, Donvil et al., 2023, Schitz et al., 9 Mar 2026, Kupervasser et al., 2011, Shu et al., 5 Dec 2025).