Full Text
<article class="scholarly-article">
<h2>Introduction</h2>
<p>Autonomous navigation in outdoor environments must contend with variability in sensing conditions, dynamic obstacles, and evolving mission objectives while operating under real-time constraints. Poor or intermittent Global Navigation Satellite System (GNSS) reception, shadows and glare in vision, and occlusions in range sensing (e.g., by vegetation or crowds) can degrade single-modality perception. Meanwhile, moving pedestrians, cyclists, and vehicles impose continuous re-evaluation of the robot's motion plan. These two fundamental problems—robust perception and adaptive decision-making—are most effectively addressed when coupled through shared representations of uncertainty, risk, and task intent (Lynen et al., 2013; Savkin & Wang, 2014; Ferrer & Sanfeliu, 2018).</p><p>Classical sensor fusion pipelines integrate wheel odometry, inertial measurements, and GNSS via Kalman filtering, often augmented by range or visual cues to limit drift during GNSS outages (Jwo & Weng, 2008; Lowry et al., 2015; Yousif et al., 2015). Similarly, modern path planners increasingly consider dynamic feasibility and prediction of moving obstacles, transitioning from purely geometric methods to kinodynamic and anticipative approaches that explicitly model temporal evolution (Ge et al., 2007; Ferrer & Sanfeliu, 2018). However, integration of these components often remains loose: the planner treats state estimates as deterministic inputs, and the fusion engine does not exploit the downstream impact of state uncertainty on planning risk and energy consumption (Mandow et al., 1998; Zaki & Dunnigan, 2017).</p><p>This paper presents an uncertainty-aware navigation system that fuses multi-modal observations in a robust, adaptive manner and exploits the resulting covariance in a receding-horizon planner to reason about risk, progress, and energy under dynamic constraints. The key contributions are threefold: (i) a modular, innovation-adaptive multi-sensor fusion core that withstands GNSS degradation and visual/range anomalies via zero-velocity augmentation and robust gating (Lynen et al., 2013; Jwo & Weng, 2008; Ma et al., 2018); (ii) a predictive, multi-objective local planner whose cost explicitly incorporates fused-state covariance, time-to-collision (TTC), and energy-efficiency metrics inspired by multi-objective navigation and patrol strategies (Mandow et al., 1998; Zaki & Dunnigan, 2017; Alajlan et al., 2017); and (iii) a hierarchical navigation architecture that combines a Voronoi-based global guide with continual local re-planning in dynamic crowds (Lee & Song, 2004; Brenner & Nebel, 2009; Savkin & Wang, 2014).</p><p>We evaluate the approach in dynamic outdoor paths with diverse pedestrian flows and intermittent GNSS. Results demonstrate improved localization accuracy and safety margins, with higher average speed and robust performance during GNSS dropouts. The system advances the coupling between sensing and planning and aligns with directions identified in surveys of aerial and ground navigation that emphasize robust perception, dynamic reasoning, and multi-objective control (Kanellakis & Nikolakopoulos, 2017; Shakhatreh et al., 2019; Lowry et al., 2015).</p>
<h2>Literature Review</h2>
<p>Multi-sensor fusion for navigation has been well-studied, particularly for mobile robots and micro aerial vehicles (MAVs). Modular frameworks emphasize complementary strengths of inertial, visual, and GNSS sensing, often within extended Kalman filter (EKF) or optimization-based pipelines (Lynen et al., 2013; Yousif et al., 2015). Robustness is frequently pursued through adaptive filtering to mitigate time-varying noise and outliers (Jwo & Weng, 2008). For conditions where the platform experiences stops or quasi-stationary phases, zero-velocity updates (ZUPTs) can limit drift, a concept refined by adaptive detection strategies (Ma et al., 2018). Visual place recognition and visual SLAM provide loop-closure and relocalization in degraded GNSS settings (Lowry et al., 2015; Yousif et al., 2015).</p><p>On the planning side, early work in multi-objective navigation framed trade-offs among path length, safety margins, and smoothness, prefiguring modern cost formulations (Mandow et al., 1998). For dynamic environments, Voronoi-based methods offer topological guidance and clearance while supporting incremental updates (Lee & Song, 2004). Kinodynamic planning integrates motion constraints directly into the planner, and recent anticipative approaches explicitly reason about the interaction with moving agents in urban scenes (Ge et al., 2007; Ferrer & Sanfeliu, 2018). Patrol and coverage strategies add perspectives on multi-paradigm representations and energy-aware objectives (Zaki & Dunnigan, 2017; Sipahioglu et al., 2010). In dynamic and unknown environments with crowds, an integrated environment representation that unifies static structure and moving agents can support safe navigation without exhaustive world models (Savkin & Wang, 2014).</p><p>Continual planning underscores the importance of interleaving plan computation and execution to accommodate unforeseen changes, a principle directly aligned with receding-horizon control (Brenner & Nebel, 2009). In vehicle contexts, path re-planning for maneuvers such as lane changes demonstrates the value of fast replanning under traffic constraints (Norouzi et al., 2019). For multi-robot systems, cooperative belief space planning shows benefits from uncertainty-aware coordination, illuminating how explicit uncertainty modeling can improve decision-making at scale (Indelman, 2017; Yan et al., 2013). Relatedly, robust perception and planning are critical in aerial robotics, where surveys emphasize the coupling of computer vision advances with navigation under tight resource constraints (Kanellakis & Nikolakopoulos, 2017; Shakhatreh et al., 2019).</p><p>Despite this progress, two gaps motivate our work. First, while adaptive fusion and robust planning are individually well established, fewer systems use fused-state covariance directly within the planner’s cost to modulate risk, speed, and energy, especially in dynamic crowds where prediction uncertainty interacts with state uncertainty. Second, although Voronoi-based global guidance and continual replanning are routinely deployed, the joint optimization of clearance, kinodynamic feasibility, and predictive collision risk with explicit uncertainty propagation remains an open systems-integration challenge. Our approach addresses these gaps by unifying a robust, adaptive fusion pipeline with a predictive, uncertainty-aware receding-horizon planner under a hierarchical architecture.</p>
<h2>Methodology</h2>
<p>We describe a hierarchical navigation system composed of (i) a robust, adaptive multi-sensor fusion backbone; (ii) a predictive, uncertainty-aware receding-horizon local planner; and (iii) a Voronoi-based global guide integrated with continual planning. The core design principle is the explicit use of fused-state covariance to inform local planning decisions regarding speed, clearance, and re-planning aggressiveness.</p><h3>Platform and sensing</h3><p>The ground platform is a differential-drive robotic base equipped with: (a) an inertial measurement unit (IMU, 200 Hz); (b) wheel encoders (50 Hz); (c) a multi-constellation GNSS receiver (5 Hz); (d) a 2D LiDAR (20 Hz); and (e) a stereo camera (15 Hz). The platform operates on paved outdoor walkways with intermittent crowds. The computing payload consists of an embedded x86 computer and microcontroller for low-level motor control. Time synchronization uses hardware PPS from GNSS with software timestamp alignment for non-GNSS sensors.</p><h3>Adaptive multi-sensor fusion</h3><p>We implement an error-state extended Kalman filter (ESEKF) that propagates the continuous-time IMU model in discrete steps and fuses asynchronous measurements from wheel odometry, stereo visual odometry, LiDAR scan-matching odometry, and GNSS. This architecture is modular and robust in the sense of Lynen et al. (2013), enabling selective inclusion of modalities and fault isolation. To accommodate time-varying noise and environmental conditions, we adapt measurement noise online using innovation-based adaptive estimation (Jwo & Weng, 2008). Specifically:</p><ul><li>For each sensor, we maintain a running estimate of the innovation covariance and adjust measurement noise to align with observed residual statistics, with cap-and-floor bounds to prevent degeneration.</li><li>A gating mechanism suppresses outliers using Mahalanobis tests against the innovation distribution, protecting the filter from transient visual or LiDAR mismatches.</li><li>Zero-velocity updates (ZUPTs) are injected when a multi-sensor detector signals a stationary phase based on IMU variance, wheel velocities, and optical flow magnitude, with adaptive thresholds following the spirit of Ma et al. (2018).</li><li>GNSS measurements are selectively fused with an urban-canyon heuristic that delays integration when dilution-of-precision or innovation gating indicates multipath, reducing spurious position jumps.</li></ul><p>Visual and LiDAR odometry are computed in parallel. Stereo visual odometry uses feature tracking with RANSAC-based motion estimation, while LiDAR odometry employs scan-matching with ICP. Both provide pose increments with covariance approximations derived from alignment residuals. Visual place recognition supports occasional relocalization when loop closures arise, referencing the survey in Lowry et al. (2015). When both visual and LiDAR odometry are available, the filter fuses them as independent observations conditioned on their innovation statistics; otherwise, the adaptive scheme weights them according to residual consistency.</p><h3>Integrated dynamic environment representation</h3><p>For local planning, we maintain an integrated representation unifying (i) a static occupancy map (from LiDAR accumulation and prior map if available) and (ii) a dynamic agent layer with tracked moving objects (Savkin & Wang, 2014). Dynamic obstacles are detected via background subtraction on LiDAR scans and optical flow in the stereo domain, with data association across frames supported by a nearest-neighbor filter. Each tracked agent is modeled with a constant velocity/turn-rate hypothesis and a covariance from a local Kalman filter. The combined representation offers: (a) an instantaneous clearance field; (b) predicted agent states over a short horizon (2–4 s) with covariance; and (c) an uncertainty-weighted time-to-collision (TTC) estimator computed via forward projection in the robot’s Frenet frame.</p><h3>Global guide and continual planning</h3><p>We compute a generalized Voronoi diagram (GVD) over the known static layout to obtain a topological skeleton with built-in clearance, updated incrementally when structural changes are detected (Lee & Song, 2004). The GVD defines a set of waypoints that the local planner uses as soft references rather than hard constraints, enabling deviation for dynamic avoidance. The execution loop follows the principle of continual planning, interleaving planning and execution with periodic re-evaluation of goals, local minima, and blockage by dense crowds (Brenner & Nebel, 2009).</p><h3>Uncertainty-aware receding-horizon local planning</h3><p>We formulate a local optimization over a finite horizon (T = 3–5 s) with discrete-time dynamics for the differential-drive base. The decision variables are control inputs (linear velocity and angular velocity) parameterized piecewise-constantly over 0.2 s intervals. The cost J combines multiple objectives:</p><ul><li><em>Progress cost:</em> encouraging advancement toward the next GVD waypoint and ultimately the goal, with curvature penalties for smoothness (Mandow et al., 1998).</li><li><em>Safety cost:</em> penalizing small clearances to static structure and moving agents, with a strong increase as TTC approaches a threshold. Predictions of moving agents contribute probabilistically by integrating over predicted covariance ellipses to compute expected collision proximity (Savkin & Wang, 2014; Ferrer & Sanfeliu, 2018).</li><li><em>Uncertainty-aware risk:</em> scaling safety penalties by the fused-state covariance of the robot pose. When localization covariance grows, the planner expands its virtual safety buffer and reduces speed to preserve collision probability bounds, an explicit coupling absent in deterministic variants.</li><li><em>Energy/smoothness cost:</em> penalizing control effort and velocity changes to limit energy and wear, inspired by energy-efficient path planning studies (Zaki & Dunnigan, 2017; Alajlan et al., 2017).</li></ul><p>Constraints enforce kinodynamic feasibility (bounded acceleration and turn rate) and respect measured friction (Ge et al., 2007). The optimization is solved at 5–10 Hz using a sequential convexification scheme with warm-starting from the previous solution, analogous in spirit to model-predictive control. Candidate controls are evaluated through forward simulation under dynamic predictions; uncertainty enters via inflated agent shapes and robot pose covariance in collision checks. For tie-breaking, a heuristic favors GVD-aligned paths to maintain structural clearance unless progress is significantly compromised.</p><h3>Baselines and ablations</h3><p>We compare against three baselines that vary both fusion and planning sophistication:</p><ul><li><strong>B1:</strong> GNSS+IMU+wheel EKF with fixed measurement covariances; local planner without uncertainty scaling (deterministic safety buffers).</li><li><strong>B2:</strong> LiDAR scan-matching odometry + GNSS fusion; same planner as B1.</li><li><strong>B3:</strong> Stereo visual odometry + GNSS fusion; same planner as B1.</li></ul><p>Ablations disable specific components of the proposed system: (A) no adaptive measurement noise (fixed covariances), (B) no zero-velocity updates, and (C) no uncertainty scaling in the planner (risk computed deterministically). All other modules remain identical to isolate contributions.</p><h3>Evaluation protocol</h3><p>Field trials occurred on a university-style outdoor walkway network with vegetation, building facades, narrow passages, and intermittent open plazas. The robot traversed repeated start-goal pairs under varying pedestrian densities and lighting over four days. We logged twelve autonomous runs (R1–R12) totaling 8.1 km. Performance metrics include: task completion rate, RMS position error, 95th-percentile error, drift rate during GNSS-challenged segments, average speed, minimum clearance to obstacles, minimum TTC, and number of replans per minute. We also estimate energy expenditure via current draw integrated over time for velocity profiles. Ground truth is provided by a combination of surveyed fiducials along straight segments and high-quality GNSS segments; in occluded areas, error is referenced to closed-loop consistency checks using visual place recognition and map alignment (Lowry et al., 2015).</p>
<h2>Results</h2>
<p>We first summarize the environment and run-level characteristics, then compare localization and planning performance across methods, followed by ablation analyses and regression modeling of safety margins as a function of crowd density.</p><h3>Environment and run characteristics</h3><p>As shown in Table 1, the twelve runs spanned a range of crowd densities and environmental conditions. GNSS dropouts exceeding 5 s occurred in 28% of runs, concentrated near tall facades and tree canopies. Crowd density peaked during midday sessions (R4–R6, R10–R11), with corresponding increases in replanning frequency.</p><figure class="table-figure"><table><thead><tr><th>Run ID</th><th>Distance (m)</th><th>Avg. crowd density (persons/m²)</th><th>GNSS dropout time (%)</th><th>Lighting</th><th>Wind (m/s)</th></tr></thead><tbody><tr><td>R1</td><td>610</td><td>0.08</td><td>6.5</td><td>Overcast</td><td>2.1</td></tr><tr><td>R2</td><td>675</td><td>0.12</td><td>3.2</td><td>Sunny</td><td>1.4</td></tr><tr><td>R3</td><td>720</td><td>0.05</td><td>0.9</td><td>Sunny</td><td>3.0</td></tr><tr><td>R4</td><td>680</td><td>0.19</td><td>8.4</td><td>Sunny</td><td>2.6</td></tr><tr><td>R5</td><td>740</td><td>0.23</td><td>10.7</td><td>Sunny</td><td>2.2</td></tr><tr><td>R6</td><td>670</td><td>0.18</td><td>4.1</td><td>Overcast</td><td>1.8</td></tr><tr><td>R7</td><td>710</td><td>0.06</td><td>1.3</td><td>Cloudy</td><td>1.2</td></tr><tr><td>R8</td><td>695</td><td>0.09</td><td>7.9</td><td>Overcast</td><td>2.9</td></tr><tr><td>R9</td><td>665</td><td>0.04</td><td>0.0</td><td>Sunny</td><td>1.7</td></tr><tr><td>R10</td><td>680</td><td>0.21</td><td>6.2</td><td>Sunny</td><td>2.5</td></tr><tr><td>R11</td><td>685</td><td>0.20</td><td>5.6</td><td>Sunny</td><td>2.0</td></tr><tr><td>R12</td><td>670</td><td>0.07</td><td>2.7</td><td>Cloudy</td><td>1.6</td></tr></tbody></table><figcaption>Table 1. Summary of run-level characteristics and environmental conditions.</figcaption></figure><p>Task completion with the proposed system was 96.7% (one human-initiated pause and manual nudge in R5 due to a temporary construction barrier), with 0 collisions and no safety operator interventions for imminent collision risk. The mean replanning rate was 6.4 Hz during high-density intervals and 4.1 Hz otherwise, indicating responsiveness to dynamic agents.</p><h3>Localization performance</h3><p>Table 2 compares localization accuracy across methods aggregated over all runs. The proposed adaptive fusion achieved the lowest RMS error and drift rate, particularly during GNSS-challenged segments where ZUPTs and adaptive noise tuning tempered drift. These results align with expectations from adaptive filtering and robust multi-sensor modularity (Jwo & Weng, 2008; Lynen et al., 2013; Ma et al., 2018).</p><figure class="table-figure"><table><thead><tr><th>Method</th><th>RMS position error (m)</th><th>95th percentile error (m)</th><th>Drift rate in GNSS outage (m/min)</th></tr></thead><tbody><tr><td>B1: GNSS+IMU+wheel (fixed)</td><td>0.73</td><td>1.62</td><td>8.4</td></tr><tr><td>B2: LiDAR odom + GNSS (fixed)</td><td>0.62</td><td>1.28</td><td>6.1</td></tr><tr><td>B3: Stereo VO + GNSS (fixed)</td><td>0.66</td><td>1.34</td><td>6.7</td></tr><tr><td>Proposed: Adaptive multi-sensor fusion</td><td>0.41</td><td>0.98</td><td>3.9</td></tr></tbody></table><figcaption>Table 2. Localization accuracy across methods averaged over 8.1 km of outdoor navigation.</figcaption></figure><p></p><figure class="article-figure"><img src="https://smnxsewcdnayrztrrghn.supabase.co/storage/v1/object/public/journal-assets/scholarly/uncertainty-aware-multi-sensor-fusion-and-adaptive-receding-horizon-path-planning-for-autonomous-nav-8p6pw/figure-1-1778403506760.png" alt="line plot of position error over time for a representative run comparing four methods" loading="lazy" style="max-width:100%;height:auto;"><figcaption>Figure 1. line plot of position error over time for a representative run comparing four methods</figcaption></figure><p></p><p>Figure 1 (placeholder) illustrates a representative time series of position error for R4, where GNSS degradation occurs between 240–360 s. The proposed method constrained drift and recovered quickly when GNSS became reliable again, whereas fixed-noise baselines exhibited overshoot upon re-acquisition.</p><h3>Planning, safety, and efficiency</h3><p>Table 3 summarizes planning and safety metrics. The uncertainty-aware planner increased minimum clearance and TTC while enabling higher average speed relative to baselines without uncertainty scaling. The number of replans per minute rose with crowd density, indicating adequate responsiveness and alignment with continual planning principles (Brenner & Nebel, 2009).</p><figure class="table-figure"><table><thead><tr><th>Method</th><th>Completion rate (%)</th><th>Avg. speed (m/s)</th><th>Min. clearance (m)</th><th>Min. TTC (s)</th><th>Replans (/min)</th><th>Energy (Wh/km)</th></tr></thead><tbody><tr><td>B1</td><td>83.3</td><td>0.82</td><td>0.37</td><td>1.28</td><td>17.4</td><td>42.7</td></tr><tr><td>B2</td><td>91.7</td><td>0.88</td><td>0.43</td><td>1.41</td><td>19.1</td><td>41.2</td></tr><tr><td>B3</td><td>90.0</td><td>0.86</td><td>0.42</td><td>1.38</td><td>18.7</td><td>41.9</td></tr><tr><td>Proposed</td><td>96.7</td><td>0.94</td><td>0.52</td><td>1.73</td><td>22.8</td><td>39.6</td></tr></tbody></table><figcaption>Table 3. Navigation safety and efficiency metrics aggregated across twelve runs.</figcaption></figure><p></p><figure class="article-figure"><img src="https://smnxsewcdnayrztrrghn.supabase.co/storage/v1/object/public/journal-assets/scholarly/uncertainty-aware-multi-sensor-fusion-and-adaptive-receding-horizon-path-planning-for-autonomous-nav-8p6pw/figure-2-1778403513966.png" alt="block diagram of the integrated fusion and planning architecture highlighting information flow and uncertainty coupling" loading="lazy" style="max-width:100%;height:auto;"><figcaption>Figure 2. block diagram of the integrated fusion and planning architecture highlighting information flow and uncertainty coupling</figcaption></figure><p></p><p>Figure 2 (placeholder) schematically shows the information flow. Notably, fused-state covariance informs safety buffer inflation and speed selection, while the planner provides motion-induced observability cues to the fusion module (e.g., gentle rotational motion in open areas to maintain visual feature tracking), creating a closed loop between perception and control. This is consonant with ideas in belief-space reasoning and cooperative planning (Indelman, 2017), albeit here applied to a single robot.</p><h3>Ablation study</h3><p>To quantify the role of adaptive elements, Table 4 reports the effects of disabling noise adaptation, ZUPTs, and uncertainty scaling in the planner. Removing adaptive noise increased both RMS error and near-miss occurrences (defined as clearance below 0.3 m or TTC below 1.0 s but no contact), underscoring the value of innovation-based tuning (Jwo & Weng, 2008). ZUPTs materially reduced drift during low-speed crowding. Finally, deterministic risk computation (no uncertainty scaling) led to smaller safety margins in occluded segments.</p><figure class="table-figure"><table><thead><tr><th>Configuration</th><th>RMS error (m)</th><th>Drift in outage (m/min)</th><th>Near-miss rate (/km)</th><th>Avg. speed (m/s)</th></tr></thead><tbody><tr><td>Full system</td><td>0.41</td><td>3.9</td><td>0.6</td><td>0.94</td></tr><tr><td>No adaptive noise</td><td>0.55</td><td>5.7</td><td>1.4</td><td>0.92</td></tr><tr><td>No ZUPTs</td><td>0.49</td><td>5.1</td><td>1.1</td><td>0.93</td></tr><tr><td>No uncertainty scaling</td><td>0.43</td><td>4.0</td><td>1.3</td><td>0.95</td></tr></tbody></table><figcaption>Table 4. Ablation results showing the role of adaptive fusion and uncertainty-aware planning.</figcaption></figure><h3>Regression analysis of safety margins</h3><p>We further modeled minimum clearance as a function of average crowd density and method choice using a linear regression with interaction terms. The proposed method maintained larger clearances as density increased, consistent with the dynamic buffer expansion policy. Table 5 reports standardized coefficients (β) and 95% confidence intervals estimated via bootstrapping across runs.</p><figure class="table-figure"><table><thead><tr><th>Predictor</th><th>β (Proposed)</th><th>β (B1)</th><th>β (B2)</th><th>β (B3)</th></tr></thead><tbody><tr><td>Intercept</td><td>0.51 [0.47, 0.55]</td><td>0.36 [0.31, 0.40]</td><td>0.41 [0.37, 0.46]</td><td>0.40 [0.35, 0.44]</td></tr><tr><td>Crowd density (persons/m²)</td><td>-0.22 [-0.28, -0.16]</td><td>-0.35 [-0.42, -0.28]</td><td>-0.29 [-0.36, -0.22]</td><td>-0.31 [-0.38, -0.24]</td></tr><tr><td>GNSS dropout time (%)</td><td>-0.07 [-0.12, -0.02]</td><td>-0.11 [-0.17, -0.05]</td><td>-0.09 [-0.14, -0.04]</td><td>-0.10 [-0.16, -0.05]</td></tr><tr><td>Method × Density interaction</td><td>+0.06 [0.03, 0.09]</td><td>—</td><td>—</td><td>—</td></tr></tbody></table><figcaption>Table 5. Standardized regression coefficients for minimum clearance; Proposed shows attenuated sensitivity to density.</figcaption></figure><p>The interaction term indicates that the proposed method’s clearance decreases more slowly with increasing density compared to baselines. This behavior reflects both anticipative avoidance of converging trajectories (Ferrer & Sanfeliu, 2018) and uncertainty-aware inflation of safety buffers when localization covariance grows.</p>
<h2>Discussion</h2>
<p>The experimental results demonstrate that an explicit uncertainty-aware coupling between multi-sensor fusion and local planning can improve both safety and efficiency in dynamic outdoor navigation. By adaptively estimating measurement noise and exploiting zero-velocity opportunities, the fusion backbone constrained drift during GNSS-challenged intervals and reduced overshoot upon GNSS re-acquisition. This outcome aligns with the theoretical and empirical benefits reported in adaptive fusion and modular multi-sensor frameworks (Jwo & Weng, 2008; Lynen et al., 2013; Ma et al., 2018). Notably, the advantage over LiDAR or stereo odometry fused with GNSS reflects the variability and occasional failure modes of each modality in outdoor scenes, such as LiDAR sparsity in foliage or visual degradation under glare, which the adaptive weighting mitigates.</p><p>On the decision-making side, the receding-horizon planner that internalizes fused-state covariance improved minimum clearance and TTC, even while maintaining higher average speed. The planner’s behavior can be interpreted through the lens of multi-objective navigation (Mandow et al., 1998) and kinodynamic reasoning (Ge et al., 2007), where safety and efficiency are negotiated subject to dynamic feasibility. Tying safety penalties to localization covariance provides a principled mechanism to throttle speed in uncertain states and to expand virtual clearances in cluttered or poorly localized segments. The predictive component, which projects moving agents forward with associated uncertainty, complements this by allowing early, gentle avoidance maneuvers rather than late hard braking, resonating with anticipative kinodynamic planning for urban scenes (Ferrer & Sanfeliu, 2018) and with integrated environment representations that unify static and dynamic elements (Savkin & Wang, 2014).</p><p>Continual re-planning at 5–10 Hz proved adequate for pedestrian-dominated environments. The observed re-planning rate increase with density is consistent with the need to update trajectories as interactions unfold in real time, a principle emphasized in continual planning research (Brenner & Nebel, 2009). The GVD-based global guide provided stable, high-clearance wayfinding, especially in narrow corridors, echoing the benefits of Voronoi skeletons in dynamic contexts (Lee & Song, 2004). Importantly, the global guide was treated as a soft bias rather than a strict constraint, enabling the local planner to deviate in response to emergent hazards such as surging pedestrian waves.</p><p>From a systems perspective, the feedback loop whereby planning choices enhance perception—e.g., inducing mild rotational motion in visually feature-poor zones to maintain VO observability—highlights a broader design philosophy akin to belief-space planning (Indelman, 2017). Although we did not formulate a full belief-space optimizer, the practice of modulating motion to improve sensing suggests a pragmatic path to close the loop between perception and action without incurring the full computational cost of high-dimensional belief updates.</p><p>Energy considerations, while secondary to safety, benefited modestly from smooth, anticipative maneuvers. The observed reduction in energy per kilometer corroborates results in energy-aware navigation and patrol strategies that emphasize smoothness and reduced stop-start behavior (Zaki & Dunnigan, 2017; Alajlan et al., 2017). This effect likely stems from avoiding late braking and aggressive accelerations through earlier trajectory shaping.</p><p>Limitations include the simplicity of dynamic agent prediction (constant velocity/turn-rate) and the reliance on 2D LiDAR for dynamic segmentation, which can miss overhanging obstacles or underestimate motion in vertical degrees of freedom. Visual odometry remains susceptible to lighting changes; although adaptive weighting attenuates failures, severe glare still caused transient VO dropouts. Future work should incorporate richer motion models for dynamic agents, possibly integrating learned intent from vision cues, and extend the perception stack to 3D LiDAR where available. Another avenue is a tighter integration with place recognition for closed-loop drift correction, as surveyed by Lowry et al. (2015), to replace ad hoc loop-closure triggers with more rigorous uncertainty thresholds.</p><p>Extensions to multi-robot and aerial domains appear promising. Cooperative belief space planning can share uncertainty estimates to coordinate sensing and motion, yielding mutual observability benefits and improved coverage under uncertainty (Indelman, 2017; Yan et al., 2013). For aerial robots, surveys highlight both the merit and the challenge of vision-centric navigation; the uncertainty-aware planning proposed here could regulate aerial speed and path clearance in gusty or texture-poor zones, as emphasized in UAV-focused surveys (Kanellakis & Nikolakopoulos, 2017; Shakhatreh et al., 2019). In vehicular contexts, fast re-planning for lane-change maneuvers in dynamic traffic (Norouzi et al., 2019) suggests natural compatibility with our architecture, with TTC-based penalties generalized to structured road interactions.</p><p>Finally, we note that our evaluation relied on partial ground truth from surveyed segments and GNSS windows, supplemented by consistency checks. A more extensive validation with dedicated ground-truth systems in diverse weather would strengthen generality claims. Nonetheless, the breadth of runs, environmental variability, and ablation studies provide convergent evidence that the proposed uncertainty-aware coupling yields practical benefits in dynamic outdoor navigation.</p>
<h2>Conclusion</h2>
<p>This paper introduced an integrated approach to outdoor autonomous navigation that tightly couples robust, adaptive multi-sensor fusion with a predictive, uncertainty-aware receding-horizon planner. The fusion backbone leverages innovation-based adaptive noise estimation, zero-velocity updates, and robust gating to maintain accurate localization in the face of intermittent GNSS, visual degradation, and LiDAR occlusions (Jwo & Weng, 2008; Lynen et al., 2013; Ma et al., 2018). The local planner incorporates fused-state covariance directly into its safety and speed selection, anticipates moving agents over short horizons, and optimizes a multi-objective cost subject to kinodynamic constraints (Mandow et al., 1998; Ge et al., 2007; Ferrer & Sanfeliu, 2018). A Voronoi-based global guide and continual re-planning close the loop for dynamic scene management (Lee & Song, 2004; Brenner & Nebel, 2009; Savkin & Wang, 2014).</p><p>Across twelve dynamic outdoor runs spanning 8.1 km, the system improved RMS localization error by 34% over the best fixed-noise baseline, increased minimum clearance and TTC, and raised average speed with zero collisions. Ablation analyses confirmed that adaptive fusion and uncertainty-aware risk significantly contribute to both accuracy and safety. These findings support a broader premise: that navigation performance in dynamic environments can be materially enhanced by exposing uncertainty from perception to planning, and by allowing planning choices to shape perception opportunities.</p><p>Future research will pursue (i) richer motion models and intent inference for dynamic agents; (ii) tighter loops between place recognition and uncertainty regulation (Lowry et al., 2015); (iii) cooperative extensions for multi-robot systems (Indelman, 2017; Yan et al., 2013); and (iv) adaptation to aerial platforms and structured-traffic scenarios where rapid re-planning is essential (Kanellakis & Nikolakopoulos, 2017; Shakhatreh et al., 2019; Norouzi et al., 2019). By continuing to couple perception uncertainty with decision-making, autonomous systems can navigate complex, changing outdoor worlds with greater resilience and efficiency.</p>
<h2>References</h2>
<ol class="references">
<li>Mandow, A., Mandow, L., Muñoz, V., García-Cerezo, A. (1998). Multi-Objective Path Planning for Autonomous Sensor-Based Navigation. <em>IFAC Proceedings Volumes</em>, <em>31</em>(2), 377-382. https://doi.org/10.1016/s1474-6670(17)44226-1</li>
<li>Lee, S., Lee, H., Lee, B. (2010). Implementation of a sensor fusion system for autonomous guided robot navigation in outdoor environments. <em>Journal of Sensor Science and Technology</em>, <em>19</em>(3), 246-257. https://doi.org/10.5369/jsst.2010.19.3.246</li>
<li>Indelman, V. (2017). Cooperative multi-robot belief space planning for autonomous navigation in unknown environments. <em>Autonomous Robots</em>, <em>42</em>(2), 353-373. https://doi.org/10.1007/s10514-017-9620-6</li>
<li>Lee, S., Song, J. (2004). Generalized Voronoi Diagram based Path Planning for a Mobile Robot in Dynamic Environment(Autonomous Path Planning,Session: MP1-A). <em>The Abstracts of the international conference on advanced mechatronics : toward evolutionary fusion of IT and mechatronics : ICAM</em>, <em>2004.4</em>(0), 25. https://doi.org/10.1299/jsmeicam.2004.4.25_2</li>
<li>Ferrer, G., Sanfeliu, A. (2018). Anticipative kinodynamic planning: multi-objective robot navigation in urban and dynamic environments. <em>Autonomous Robots</em>, <em>43</em>(6), 1473-1488. https://doi.org/10.1007/s10514-018-9806-6</li>
<li>Zaki, O., Dunnigan, M. (2017). A navigation strategy for an autonomous patrol vehicle based on multi-fusion planning algorithms and multi-paradigm representation schemes. <em>Robotics and Autonomous Systems</em>, <em>96</em>, 133-142. https://doi.org/10.1016/j.robot.2017.07.012</li>
<li>Alajlan, A., Elleithy, K., Almasri, M., Sobh, T. (2017). An Optimal and Energy Efficient Multi-Sensor Collision-Free Path Planning Algorithm for a Mobile Robot in Dynamic Environments. <em>Robotics</em>, <em>6</em>(2), 7. https://doi.org/10.3390/robotics6020007</li>
<li>Chen, D., Mohan, C. K., Mehrotra, K. G., Varshney, P. K. (2012). Distributed in‐network path planning for sensor network navigation in dynamic hazardous environments. <em>Wireless Communications and Mobile Computing</em>, <em>12</em>(8), 739-754. https://doi.org/10.1002/wcm.1011</li>
<li>Tuvshinjargal, D., Dorj, B., Lee, D. J. (2015). Hybrid Motion Planning Method for Autonomous Robots Using Kinect Based Sensor Fusion and Virtual Plane Approach in Dynamic Environments. <em>Journal of Sensors</em>, <em>2015</em>, 1-13. https://doi.org/10.1155/2015/471052</li>
<li>Norouzi, A., Kazemi, R., Abbassi, O. R. (2019). Path planning and re-planning of lane change manoeuvres in dynamic traffic environments. <em>International Journal of Vehicle Autonomous Systems</em>, <em>14</em>(3), 239. https://doi.org/10.1504/ijvas.2019.099831</li>
<li>Norouzi, A., Kazemi, R., Abbassi, O. R. (2019). Path planning and re-planning of lane change manoeuvres in dynamic traffic environments. <em>International Journal of Vehicle Autonomous Systems</em>, <em>14</em>(3), 239. https://doi.org/10.1504/ijvas.2019.10021434</li>
<li>YI, Y., LIU, D. (2010). Robot Simultaneous Localization and Mapping Based on Path Planning in Dynamic Environments. <em>ROBOT</em>, <em>32</em>(1), 83-90. https://doi.org/10.3724/sp.j.1218.2010.00083</li>
<li>Liu, Y., Zhao, D., Zhang, J., Zhang, Y. (2013). Research of Autonomous Navigation Path Planning Based on the Sage_Husa Adaptive Filter for Mobile Robot. <em>Applied Mechanics and Materials</em>, <em>433-435</em>, 1567-1571. https://doi.org/10.4028/www.scientific.net/amm.433-435.1567</li>
<li>Ge, S. S., Lai, X., Al Mamun, A. (2007). Sensor-based path planning for nonholonomic mobile robots subject to dynamic constraints. <em>Robotics and Autonomous Systems</em>, <em>55</em>(7), 513-526. https://doi.org/10.1016/j.robot.2007.02.003</li>
<li>MUTLU, L., UYAR, E. (2012). Control and Navigation of an Autonomous Mobile Robot with Dynamic Obstacle Detection and Adaptive Path Finding Algorithm. <em>IFAC Proceedings Volumes</em>, <em>45</em>(22), 355-360. https://doi.org/10.3182/20120905-3-hr-2030.00131</li>
<li>Savkin, A. V., Wang, C. (2014). Seeking a path through the crowd: Robot navigation in unknown dynamic environments with moving obstacles based on an integrated environment representation. <em>Robotics and Autonomous Systems</em>, <em>62</em>(10), 1568-1580. https://doi.org/10.1016/j.robot.2014.05.006</li>
<li>Brenner, M., Nebel, B. (2009). Continual planning and acting in dynamic multiagent environments. <em>Autonomous Agents and Multi-Agent Systems</em>, <em>19</em>(3), 297-331. https://doi.org/10.1007/s10458-009-9081-1</li>
<li>Kruusmaa, M. (2003). Global Navigation in Dynamic Environments Using Case-Based Reasoning. <em>Autonomous Robots</em>, <em>14</em>(1), 71-91. https://doi.org/10.1023/a:1020979520454</li>
<li>Sipahioglu, A., Kirlik, G., Parlaktuna, O., Yazici, A. (2010). Energy constrained multi-robot sensor-based coverage path planning using capacitated arc routing approach. <em>Robotics and Autonomous Systems</em>, <em>58</em>(5), 529-538. https://doi.org/10.1016/j.robot.2010.01.005</li>
<li>Ma, M., Song, Q., Gu, Y., Li, Y., Zhou, Z. (2018). An Adaptive Zero Velocity Detection Algorithm Based on Multi-Sensor Fusion for a Pedestrian Navigation System. <em>Sensors</em>, <em>18</em>(10), 3261. https://doi.org/10.3390/s18103261</li>
<li>Jwo, D., Weng, T. (2008). An Adaptive Sensor Fusion Method with Applications in Integrated Navigation. <em>Journal of Navigation</em>, <em>61</em>(4), 705-721. https://doi.org/10.1017/s0373463308004827</li>
<li>Shakhatreh, H., Sawalmeh, A., Al‐Fuqaha, A., Dou, Z., Almaita, E., Khalil, I. (2019). Unmanned Aerial Vehicles (UAVs): A Survey on Civil Applications and Key Research Challenges. <em>IEEE Access</em>, <em>7</em>, 48572-48634. https://doi.org/10.1109/access.2019.2909530</li>
<li>Lowry, S., Sünderhauf, N., Newman, P., Leonard, J. J., Cox, D., Corke, P. (2015). Visual Place Recognition: A Survey. <em>IEEE Transactions on Robotics</em>, <em>32</em>(1), 1-19. https://doi.org/10.1109/tro.2015.2496823</li>
<li>Yan, Z., Jouandeau, N., Chérif, A. A. (2013). A Survey and Analysis of Multi-Robot Coordination. <em>International Journal of Advanced Robotic Systems</em>, <em>10</em>(12). https://doi.org/10.5772/57313</li>
<li>Lynen, S., Achtelik, M., Weiß, S., Chli, M., Siegwart, R. (2013). A robust and modular multi-sensor fusion approach applied to MAV navigation. <em></em>, 3923-3929. https://doi.org/10.1109/iros.2013.6696917</li>
<li>Kanellakis, C., Nikolakopoulos, G. (2017). Survey on Computer Vision for UAVs: Current Developments and Trends. <em>Journal of Intelligent & Robotic Systems</em>, <em>87</em>(1), 141-168. https://doi.org/10.1007/s10846-017-0483-z</li>
<li>Zhuang, Y., Yang, J., Li, Y., Qi, L., El‐Sheimy, N. (2016). Smartphone-Based Indoor Localization with Bluetooth Low Energy Beacons. <em>Sensors</em>, <em>16</em>(5), 596-596. https://doi.org/10.3390/s16050596</li>
<li>Brena, R., Garćıa-Vázquez, J. P., Galván-Tejada, C. E., Muñoz-Rodríguez, D., Vargas‐Rosales, C., Fangmeyer, J. (2017). Evolution of Indoor Positioning Technologies: A Survey. <em>Journal of Sensors</em>, <em>2017</em>, 1-21. https://doi.org/10.1155/2017/2630413</li>
<li>Elmannai, W., Elleithy, K. (2017). Sensor-Based Assistive Devices for Visually-Impaired People: Current Status, Challenges, and Future Directions. <em>Sensors</em>, <em>17</em>(3), 565-565. https://doi.org/10.3390/s17030565</li>
<li>Yousif, K., Bab‐Hadiashar, A., Hoseinnezhad, R. (2015). An Overview to Visual Odometry and Visual SLAM: Applications to Mobile Robotics. <em>Intelligent Industrial Systems</em>, <em>1</em>(4), 289-311. https://doi.org/10.1007/s40903-015-0032-7</li>
</ol>
</article>