TRACE: Traversability-aware Reactive Navigation via VLM-driven Adaptive Constraints Estimation

1. INTRODUCTION

• Forest terrain varies in surface conditions and obstacles, while appearance or semantic labels alone may not indicate whether the ground is physically traversable.Dynamic lighting, seasonal variation, and vegetation occlusions can also degrade segmentation and make terrain interpretation inconsistent.Source: §1 · p. 1
• Humans use worn trails, clearings, and subtle traces of prior movement as cues to accessibility; VLMs can associate visual patterns with concepts such as trails and obstacles.The paper motivates these cues as a human-inspired alternative to relying only on semantic categories or geometric thresholds.Source: §1 · pp. 1–2

1.1 Contributions

• TRACE combines positive- and negative-prompt VLM confidence maps through evidential fusion to create a dense traversability map, then derives corridor boundaries for NMPC soft barrier constraints.The overview describes projecting the map into 3D, percentile-binning annotated points, and fitting splines or regularized polynomials to the boundaries.Source: §1.1 · p. 2, Figure 1 · p. 2
• TRACE is evaluated in real-world forest environments in northern Sweden using a Husky mobile robot, with field tests reporting safe navigation across cluttered, deformable, and visually ambiguous terrain.The contributions passage describes the results as demonstrating generalization across varying environments.Source: §1.1 · p. 2

Classical Traversability Analysis and Learning-Based Approaches

• Geometry-based methods estimate traversability from terrain morphology but can underperform on vegetated or deformable terrain; semantic segmentation remains sensitive to illumination, occlusion, and domain shifts.Multimodal systems combine LiDAR and camera data or geometric and visual cues.Source: §2 · p. 2
• Learning-based methods address generalization and dynamic terrain, but the paper says most classical and learning-based approaches still rely on explicit geometric or semantic labels.The authors identify this reliance as limiting inference of contextual or affordance-level traversability cues.Source: §2 · p. 2

VLM-based and Context-Aware Traversability

• Recent work applies VLMs to context-aware navigation, outdoor traversal reasoning, obstacle interaction, mapless navigation, zero-shot traversability, and human-in-the-loop off-road adaptation.The cited examples span indoor and outdoor settings and diverse navigation tasks.Source: §2 · pp. 2–3
• TRACE explores human-inspired VLM inference from implicit cues such as loosely structured trails and surface traces for context-aware forest navigation beyond purely geometric or semantic methods.The paper frames this as a way to reason about traversability from environmental affordances.Source: §2 · p. 3

3. PROBLEM FORMULATION

• The navigation objective is to dynamically infer a traversable corridor with VLMs and find an optimal strategy for a robot operating in a forest.The left and right boundaries are represented as bounding polynomials and imposed as NMPC barrier constraints.Source: §3 · p. 3
• The robot model uses position (x, y) and heading θ as states, with commanded linear velocity v and angular velocity ω as control inputs.The model is attributed to Wang et al. (2023).Source: §3 · p. 3
• The traversable region is approximated by two degree-d polynomials representing its left and right boundaries, with coefficients C_L and C_R.The admissible lateral position lies between the boundaries after applying a positive safety margin m.Source: §3 · p. 3

VLM-based Traversability Scoring

• Positive and negative prompt sets provide pixelwise evidence that evidential fusion uses to estimate traversability and uncertainty.Prompts describe traversable terrain and obstacles; the fused map assigns higher values to more likely traversable regions.Source: §4 · p. 3

• Depth points are reprojected into the image to receive traversability scores, forming an augmented point cloud of positions and scores.Source: §4 · p. 3

Percentile-based Edge Extraction

• Points below a confidence threshold are removed, and the remaining lateral coordinates are binned along the forward axis to estimate corridor edges by percentiles.The lower and upper percentiles are 5% and 95%.Source: §4 · p. 3

Polynomial Edge Fitting

• Boundary samples are smoothed and fitted with regularized polynomials or splines to produce continuous left and right corridor edges.The fitting hierarchy uses cubic smoothing splines, weighted Chebyshev ridge regression if spline fitting is unstable, and weighted monomial ridge regression as a fallback.Source: §4 · pp. 3–4

5.1 Performance Comparison

• On GOOSE samples, TRACE produces sharper, more localized traversability activations than Gasparino et al. (2022), aligning more closely with drivable regions across varying conditions.The comparison used only the bottom half of images to match the baseline’s lower field of view.Source: §5.1 · p. 4
• The Stable terrain class has the highest IoU against Guan et al. (2022b) labels, followed by the Granular class.The authors interpret this ordering as consistency between TRACE estimates and the terrain navigability annotations.Source: Table 1 · p. 4, §5.1 · p. 4
• TRACE’s traversable-area tilt profile is smoother and more consistent than the LiDAR-only baseline, whose tilt-angle standard deviation is higher.The difference is especially relevant in cluttered or visually ambiguous environments, where geometric cues become noisy.Source: Figure 3 · p. 4, §5.1 · pp. 4–5
• In the reported examples, TRACE rejects a sloped rock as unsafe and infers a continuous path through dense grass where the LiDAR baseline fails to delineate boundaries.The comparison illustrates boundary estimation in cases involving slopes and vegetation.Source: Figure 4 · p. 5, §5.1 · p. 5

5.2 Field Test

• On a northern Sweden Husky robot, TRACE followed loosely defined forest roads and hiking trails across tall grass, overgrowth, rocks, vegetation, slopes, and meandering passages.The field trials are described as tracing loosely defined signs of previously traversed paths.Source: §5.2 · p. 5

6. CONCLUSION

• TRACE integrates vision–language perception with adaptive contour estimation for reactive motion planning in forest environments.Source: §6 · pp. 5–6
• VLM-driven traversability maps, percentile-based boundary extraction, and polynomial contour fitting support smooth, stable NMPC integration in difficult terrain.The conclusion describes the target terrains as cluttered, deformable, and visually ambiguous.Source: §6 · p. 6
• Real-world experiments confirmed adaptation to different environmental conditions and high spatial consistency with stable motion in unstructured trails.Future work will extend evaluation to more environments and robotic platforms, including legged and humanoid robots.Source: §6 · p. 6