Autonomous Hiking Trail Navigation Via Semantic Segmentation and Geometric Analysis

Authors: Camndon Reed, Christopher Tatsch, Jason N. Gross, Yu Gu · IROS 2025

Licences

Open Access
Paper: CC BY 4.0

Everything below draws on the full paper under its CC BY 4.0 licence; its figure is credited to the authors.

TL;DR

Camera semantics and LiDAR geometry are fused into one traversability map; an RRT* waypoint planner then keeps a Husky robot on the trail but sends it off-trail around hazards.

Why it matters

The authors list trail monitoring and maintenance, guided tours, wildlife tracking, ecosystem assessment and search and rescue as uses for trail-capable robots. The simulation sweep makes the key trade-off explicit: more semantic weight keeps the robot on the trail more, which the authors link to lower erosion and habitat disturbance, at the cost of longer routes.

Problem

Hiking trails are unstructured and change with weather, vegetation, erosion and human traffic, which makes them hard for autonomous ground vehicles. LiDAR-only geometric methods can miss small trail hazards or treat tall grass as an obstacle, camera-only semantic methods suffer in poor weather and low visibility, and 2.5D fused grid maps store one value per cell and miss overhanging or multi-level structure. A trail robot also has to balance staying on the trail against leaving it to avoid hazards or take safe shortcuts.

Proposed solution

A YOLOv8 segmentation model trained on 1,250 hand-labelled trail images (8 classes) labels the ZED 2i RGB stream; the masks are projected onto stereo-depth point clouds, which are filtered (0.5–6.0 m range, 0.1 m voxels, statistical outlier removal) and paired by dual k-d tree with a LiDAR geometric terrain analysis. Each point's cost is C = w·(semantic cost) + (1 − w)·(geometric cost), and fused clouds are registered into a global traversability map with odometry and ICP. An RRT* global planner samples candidate waypoints from the traversability distribution instead of uniformly and hands them to the FALCO local planner.

Key insights

Geometry-only traversability (w = 0) gave the worst results, and weights below 0.5 performed worse on all three metrics (time, distance, share of driving on trail).
Between w = 0.5 and w = 1, more semantic weight raised the share of driving on trail but lengthened traversal time and distance, a clear adherence-versus-efficiency trade-off.
Biasing RRT* sampling by the traversability map's probability distribution steers tree growth toward traversable ground rather than uniform random samples.
Filtering out points above the robot's height and outside the stereo camera's reliable 0.5–6.0 m depth range reduces noise from false depth and dense vegetation.

Main results

The YOLOv8 segmentation model reached 75.8% mAP across all classes and 96.7% mAP on the trail class.
Simulated Path 1: share of driving on trail rose from 24.93% at w = 0 to 67.47% at w = 1; w = 0.5 gave the shortest time and distance (114.14 s, 140.8 m).
Simulated Path 2: share on trail rose from 18.768% at w = 0 to 64.65% at w = 1, but w = 1 took 193.64 s and 225.1 m versus 108.47 s and 146.2 m at w = 0.5.
Field test on the West Virginia University Core Arboretum trail with w = 0.75: the robot identified impassable objects, detoured through grass and returned to the trail once a clear path was available.

Limitations

The authors call for a larger segmentation dataset to improve accuracy on under-represented classes, and for data from other seasons and trails to improve generalisation.
They state that the waypoint selector's balance between trail adherence and off-trail detours still needs optimisation.
The field demonstration covers one trail section and is reported qualitatively; the traversability map was built during a teleoperated pass before waypoints were planned.
Semantic mapping relies on stereo depth that the authors treat as accurate only between 0.5 m and 6.0 m.
The fusion weight w is a fixed, hand-set parameter chosen from two simulated paths (0.75 in the field).

Field impact

The paper was an IROS 2025 award finalist (program session TuAT1) and has 4 citations. The authors release a semantically labelled hiking-trail image dataset on IEEE DataPort and a Gazebo hiking-trail simulation environment intended as a benchmark.

Abstract

Natural environments pose significant challenges for autonomous robot navigation, particularly due to their unstructured and ever-changing nature. Hiking trails, with their dynamic conditions influenced by weather, vegetation, and human traffic, represent one such challenge. This work introduces a novel approach to autonomous hiking trail navigation that balances trail adherence with the flexibility to adapt to off-trail routes when necessary. The solution is a Traversability Analysis module that integrates semantic data from camera images with geometric information from LiDAR to create a comprehensive understanding of the surrounding terrain. A planner uses this traversability map to navigate safely, adhering to trails while allowing off-trail movement when necessary to avoid on-trail hazards or for safe off-trail shortcuts. The method is evaluated through simulation to determine the balance between semantic and geometric information in traversability estimation. These simulations tested various weights to assess their impact on navigation performance across different trail scenarios. Weights were then validated through field tests at the West Virginia University Core Arboretum, demonstrating the method's effectiveness in a real-world environment.

Open Source Robotics