TRACE: Traversability-aware Reactive Navigation via VLM-driven Adaptive Constraints Estimation
Authors: Mario A. V. Saucedo, Midhun Edatharam Kunnath, Elias Small, Akash Patel, Shruti Kotpalliwar, Christoforos Kanellakis, George Nikolakopoulos
Organizations: Robotics and AI Team, Lule˚a University of Technology, Lule˚a, SE-97187, Sweden. · Korea Advanced Institute of Science & Technology (KAIST), Daejeon, 34141, South Korea.
This paper presents TRACE, a traversability-aware reactive navigation framework designed for the autonomous operation of ground robots in unstructured forest environments. Unlike traditional approaches, TRACE leverages VLM-based confidence maps from positive and negative natural-language prompt sets and combines them through evidential fusion to produce a dense traversability map that prioritizes the terrain based on signs of previously traversed paths. Traversability-annotated points are then extracted and used to derive the left and right boundaries of the traversable path via percentile binning followed by spline or regularized polynomial fitting. These boundaries are then incorporated as soft barrier constraints within an NMPC formulation, enabling safe and adaptive path following in complex natural environments. Rigorous experimental trials carried out in real-life forest environments demonstrate TRACE ability for smooth navigation in cluttered, deformable, and visually ambiguous terrains.
Figures & tables
Figure 1: Framework Overview. Block diagram of the proposed framework, where VLM-based confidence maps from positive and negative prompt sets are combined through evidential fusion to produce a dense traversability map. The map is then projected onto 3D to obtain traversability-annotated points, which are later used to extract the left and right traversable boundaries based on percentile binning and spline/regularized polynomial fitting. Finally, the resulting boundaries are incorporated as soft barrier constraints within the NMPC formulation.
Figure 2: Traversability Maps. Visualizations of the estimated traversability maps on different challenging samples from the GOOSE dataset ( Mortimer et al. (2024) ), illustrating the comparison between the baseline’s (i.e., middle row) ( Gasparino et al. (2022) ) and the proposed method’s (i.e., bottom row) ability to adapt to different environmental conditions.
Label
Stable
Granular
Poor foothold
High resistance
Obstacle
IoU
0.672421
0.334350
0.070665
0.007723
0.001812
Table 1: IoU for the estimated terrain traversability on the Guan et al. (2022b) labels.
Figure 3: Normal Tilt Angle. Average surface-normal tilt angle for the traversable area estimated by the proposed semantic-based method compared to the LiDAR-only baseline.
Figure 4: Path Boundaries. Comparison of path boundary estimation between LiDAR-only baseline and the proposed approach, where the baseline misclassifies certain sloped or vegetated areas as traversable, while TRACE correctly identifies them as unsafe and extracts smooth, coherent boundaries around the truly navigable regions.
Figure 5: Exp. 1. Forest road obstructed by overgrow vegetation.
Figure 6: Exp. 2. Narrow hiking trail surrounded by rocks and vegetation.
Field Robotics Engineering and Science Hub (FRESH), Illinois Autonomous Farm, University of Illinois at Urbana-Champaign (UIUC), IL. · Mobile Robotics Group, University of S˜ao Paulo, S˜ao Carlos, SP, Brazil.