---
title: 'DuLoc: Dual-Layer LiDAR Localization'
url: https://www.emergentmind.com/topics/duloc
type: topic
---

# DuLoc: Dual-Layer LiDAR Localization

Searching arXiv for "DuLoc" and closely related entries to disambiguate the topic and ground the article in the relevant paper(s).
DuLoc denotes a LiDAR-based localization method for autonomous systems introduced in "DuLoc: Life-Long Dual-Layer Localization in Changing and Dynamic Expansive Scenarios" [2507.23660]. It is designed for large-scale changing outdoor environments in which conventional point-cloud registration against an offline map alone can lose robustness under long-term environmental change, producing localization drift and degraded reliability. DuLoc addresses this by tightly coupling a LiDAR-IMU odometry backbone, an offline static prior map that provides global constraints, and a rolling local dynamic map built online, while also incorporating a constant-velocity motion model to mitigate outlier noise and brief sensor faults [2507.23660].

## 1. Problem setting and operational scope

DuLoc is formulated for localization in changing and dynamic expansive scenarios, with particular emphasis on unbounded environments and long-duration deployment. The motivating observation is that methods relying solely on offline maps often have limited robustness against long-term environmental changes, whereas pure odometric pipelines can accumulate drift. DuLoc therefore targets a regime in which repeatability, accuracy, and environmental adaptability must be optimized simultaneously rather than in isolation [2507.23660].

The deployment context reported for DuLoc is an operational port area of approximately \(1\) million \(\text{m}^2\), with dynamic elements including cranes, containers, and vehicles, as well as multiple GPS-denied zones. The evaluation platform is an Industrial Intelligent Guided Vehicle (IGV) of size \(15\,\text{m} \times 3\,\text{m} \times 1.7\,\text{m}\), equipped with two 16-line Lidars at \(10\,\text{Hz}\) each, a 6-axis IMU at \(100\,\text{Hz}\), and RTK ground truth at \(50\,\text{Hz}\). The reported operational corpus comprises \(2{,}856\) hours of real runs across \(32\) different IGVs and approximately \(6\,\text{km}\) test paths under varied lighting, weather, and dynamic loads [2507.23660].

This operating envelope suggests that DuLoc is not aimed at a narrow benchmark setting. A plausible implication is that its principal contribution lies in maintaining localization validity under prolonged environmental nonstationarity rather than merely improving single-run registration accuracy.

## 2. Dual-layer estimation architecture

The core architectural idea is a dual-layer localization framework embedded in a single iterated Error-State Kalman Filter (iESKF). The three fused components are: a tightly-coupled LiDAR-IMU odometry backbone, an offline static prior map for global constraints, and a rolling local dynamic map for recent observations. The combination is intended to provide long-term consistency through global-map registration and short-term robustness through local-map registration and inertial prediction [2507.23660].

At time step \(i\), DuLoc maintains a \(30\)-dimensional state vector
\[
\mathbf{x}_i =
\bigl[\,^{W}\!R_{I_i},\; ^{W}\!p_{I_i},\; ^{W}\!v_{I_i},\;
\omega_i,\; a_i,\;
b_{\omega_i},\; b_{a_i},\;
^{W}\!g,\;
^{I}\!R_{L},\; ^{I}\!p_{L}\bigr].
\]
Here \(^{W}\!R_{I_i},\,^{W}\!p_{I_i},\,^{W}\!v_{I_i}\) are the IMU pose and velocity in the world frame; \(\omega_i\) and \(a_i\) are measured angular velocity and acceleration; \(b_{\omega_i}\) and \(b_{a_i}\) are IMU bias states; \(^{W}\!g\) is gravity; and \((^{I}\!R_L,^{I}\!p_L)\) are the known LiDAR-IMU extrinsics [2507.23660].

The IMU-driven motion model is expressed through a continuous-time forward model \(f(\mathbf{x},\mathbf{w})\) with zero-mean Gaussian noise \(\mathbf{w}\), yielding the discrete propagation
\[
\mathbf{x}_{i+1}
= \mathbf{x}_i \;\boxplus\;\bigl(\Delta t\,f(\mathbf{x}_i,\mathbf{w}_i)\bigr).
\]
To handle IMU dropouts or severe sensor faults, DuLoc falls back to a \(100\,\text{Hz}\) constant-velocity model by setting \(\mathbf{w}=0\) and retaining only the current linear and angular velocities:
\[
\mathbf{x}_{i+1}
= \mathbf{x}_i
\;\boxplus\;\bigl(\Delta t\,f(\mathbf{x}_i,\mathbf{0})\bigr).
\]
This fallback is explicitly presented as a divergence-avoidance mechanism during brief data interruptions rather than as a replacement for inertial propagation [2507.23660].

## 3. Fusion pipeline and registration model

DuLoc’s estimation loop uses an iESKF with asynchronous prediction and correction. Prediction integrates IMU data, or the constant-velocity substitute when necessary, at up to \(100\,\text{Hz}\), propagating both the mean state \(\mathbf{x}_i\) and covariance \(\mathbf{P}_i\). Correction is triggered by incoming LiDAR sweeps or IMU measurements. LiDAR point-to-plane residuals against both the global and local maps are inserted directly as measurement updates, while IMU biases and gravity are corrected by treating raw IMU readings as pseudo-measurements [2507.23660].

For a cloud frame \(i\), planar features \(^{L_i}\!p_j\) with normals \(^{L_i}\!n_j\) are extracted. If a feature is associated with a planar patch in map \(\mathcal{M}\), DuLoc defines the residual
\[
r_j(\mathbf{x}_i)
= \bigl(^{\mathcal{M}}\!u_j\bigr)^\top
\Bigl(
{}^{\mathcal{M}}\!T_{I_i}\,{}^{I_i}\!T_{L_i}\,
\bigl(^{L_i}\!p_j +\,^{L_i}\!n_j\bigr)
-\;^{\mathcal{M}}\!q_j
\Bigr),
\]
where \(^{\mathcal{M}}\!u_j\) is the map plane normal and \(^{\mathcal{M}}\!q_j\) is any in-plane point. The distinctive step is a dual-map residual constraint per point:
\[
0 = r^{\rm global}_j(\mathbf{x}_i)\;+\;r^{\rm local}_j(\mathbf{x}_i).
\]
This couples alignment to the offline prior map \(\mathcal{M}_{\rm global}\) and to the online dynamic map \(\mathcal{M}_{\rm local}\) within the same filter update [2507.23660].

The computational side is also explicit. Building on FAST-LIO’s state-dependent gain trick, DuLoc projects the large point-cloud measurement dimension down to the state dimension when computing the Kalman gain, preserving \(O(n_{\rm state}^2)\) complexity instead of \(O(n_{\rm obs}^3)\). This detail is central to understanding why dual-map updates remain feasible at real-time rates despite operating on dense geometric observations [2507.23660].

## 4. Prior-map construction and local dynamic map management

DuLoc separates map management into an offline static layer and an online dynamic layer. The prior map is constructed offline from long runs using Fast-LIO2 combined with a high-grade INS to align sweeps over approximately \(1\) million \(\text{m}^2\). It is then post-processed to statistically remove dynamic objects, or “ghosting,” such that only static and repeatable structures such as ground and pillars remain [2507.23660].

The online local dynamic map is maintained by inserting distortion-corrected LiDAR points from each new sweep into an incremental k-d tree, denoted ikd-Tree. The local map is periodically pruned according to three criteria: points that are too old, points beyond a fixed radius from the current pose, and points tagged as dynamic outliers. In the reported design, limiting the local map to a spatial radius of roughly \(50\)–\(100\,\text{m}\) around the vehicle bounds memory usage in unbounded scenes while keeping nearest-neighbor lookup above \(10\,\text{kHz}\) [2507.23660].

This map strategy gives DuLoc its “dual-layer” character in an operational sense. The prior layer stabilizes global localization against drift, while the local layer absorbs recent, scene-specific structure and thereby accommodates environmental changes that are absent from the offline map. The paper’s discussion states this more generally: as long as either the prior map or the local map retains sufficient structure, pose updates remain robust [2507.23660].

## 5. Experimental evaluation and reported performance

The reported evaluation uses Absolute Trajectory Error (ATE) computed with the evo toolkit, lateral and longitudinal drift, and availability, defined as the percentage of time localization is valid. The experiments cover large-scale changing outdoor environments and compare DuLoc against DLL, FAST-LIO2, Light-LOAM, and Point-LIO [2507.23660].

On the representative run corresponding to Fig. 4(d), the reported errors in meters are as follows. DLL yields Max ATE \(451.69\), Mean ATE \(207.60\), Max Lat \(231.12\), Mean Lat \(63.02\), Max Long \(444.84\), and Mean Long \(100.32\). FAST-LIO2 yields \(8.23\), \(2.75\), \(8.05\), \(1.88\), \(2.85\), and \(0.27\). Light-LOAM yields \(182.99\), \(149.81\), \(96.16\), \(68.09\), \(176.48\), and \(130.29\). Point-LIO yields \(55.52\), \(3.59\), \(5.64\), \(1.56\), \(55.24\), and \(1.13\). DuLoc yields \(0.74\), \(0.17\), \(0.24\), \(0.03\), \(0.70\), and \(0.03\) [2507.23660].

For the runs corresponding to Fig. 4(g) and Fig. 4(j), DuLoc reports Max ATE / Mean ATE of \(0.93 / 0.20\) and \(0.55 / 0.26\), respectively; the corresponding lateral errors are \(0.90 / 0.07\) and \(0.51 / 0.12\), and the longitudinal errors are \(0.53 / 0.03\) and \(0.47 / 0.08\) [2507.23660].

Across the study, DuLoc is reported to achieve centimeter-level mean ATE, with mean \(\approx 0.2\,\text{m}\), and \(99.98\%\) availability. Competing map-based methods such as DLL and Light-LOAM are reported to catastrophically fail within \(30\,\text{min}\) in dynamic port runs. Runtime on an Intel i7-8750H with \(32\,\text{GB}\) RAM and a GTX 1050Ti is given as real-time operation with IMU/CV prediction at \(100\,\text{Hz}\), LiDAR update at \(10\,\text{Hz}\), and map queries above \(15\,\text{kHz}\) [2507.23660].

These results support the paper’s claim that the benefit of the method is not limited to nominal accuracy. The combination of low mean ATE and very high availability indicates that DuLoc is optimized for sustained operational validity under changing outdoor conditions.

## 6. Limitations, extensions, and terminological ambiguity

The reported limitations are concrete. Extreme scene changes beyond the static prior map, such as newly erected gantry cranes, can reduce global-map matches until the local map adapts. Very aggressive maneuvers with high IMU saturation can degrade the constant-velocity fallback. The main computational cost is LiDAR point matching against two maps, mitigated through the ikd-Tree and the state-dimension Kalman-gain computation. Proposed extensions include semantic segmentation to identify dynamic vehicles and cranes, hierarchical multi-resolution prior maps, and learnable motion priors for the fallback model, such as turn-rate limits on IGVs [2507.23660].

A separate issue is terminological ambiguity. The label “DuLoc” also appears in an exposition of the Prolog framework introduced in "Disjunctive Delimited Control," where it denotes a system for capturing both conjunctive and disjunctive continuations using `shift/1` and `reset/3`, rather than a localization method [2009.04909]. In that usage, `reset(Pattern,Goal,Result)` always succeeds and returns one of `failure`, `success(PatternCopy,DisjCont)`, or `shift(Term,ConjCont,PatternCopy,DisjCont)`, and the framework is presented as enabling library-level implementations of `findall`, cut, branch-and-bound optimisation, and probabilistic programming [2009.04909].

For robotics literature, however, DuLoc most prominently refers to the dual-layer LiDAR localization framework of [2507.23660]. Its defining characteristic is the tight single-filter integration of high-frequency IMU or constant-velocity propagation, registration against a cleaned static prior map, and registration against a rolling dynamic local map, with the explicit goal of life-long localization in changing and expansive real-world scenes [2507.23660].

Source: https://www.emergentmind.com/topics/duloc