Abstract
We present a novel system, AdVENTR, for autonomous robot navigation in unstructured outdoor environments consisting of uneven and vegetated terrains. Our approach is general and enables both wheeled and legged robots to handle outdoor terrain complexity, including unevenness, surface properties like poor traction, granularity, and obstacle stiffness. Using data from RGB cameras, 3D LiDAR, IMU, and robot odometry, we employ efficient learning-based perception and planning algorithms capable of execution on edge computing hardware. A core component of our system is a scene-aware switching method that dynamically selects between multiple perception algorithms based on the immediate environment. We demonstrate AdVENTR's performance on Husky and Spot robots across various sloped, rocky, muddy, and densely vegetated terrains.
Introduction
Navigating unstructured outdoor environments presents a "perfect storm" of robotics challenges. Unlike controlled indoor settings with flat floors and rigid walls, outdoor terrains are characterized by:
- Non-rigid obstacles: Tall grass or brush may appear as solid walls to traditional laser scanners, yet they are often traversable.
- Granular and slippery surfaces: Sand, mud, and mulch introduce wheel slip and instability that standard geometric planners fail to predict.
- Computational constraints: Running high-fidelity deep learning models for every possible terrain type simultaneously is often too resource-intensive for a robot's onboard edge computer.
Existing navigation systems often fail because they treat all obstacles as rigid or assume uniform traction. AdVENTR (Adaptive Vegetation, Elevation, aNd Terrain-aware Robot) addresses these issues by creating a "scene-aware" system that identifies the specific challenges of its immediate surroundings and adjusts its perception strategy accordingly.
System Architecture
The AdVENTR framework decomposes navigation into a modular pipeline designed for efficiency and robustness.
Scene-aware Switching
The core innovation of AdVENTR is its scene-aware switching mechanism. Instead of running every perception algorithm concurrently, the system uses lightweight metrics to detect specific environmental features:
- Terrain Unevenness: Detected via elevation gradients.
- Surface Properties: Identified through PCA analysis of IMU and joint data.
- Vegetation: Detected via LiDAR intensity.
Based on these detections, the system activates only the necessary specialized perception modules, optimizing computational load.
Specialized Perception Modules
- TERP: Analyzes elevation maps to identify paths with minimal slope and maximum stability.
- TerraPN/ProNAV: Utilizes self-supervised learning to correlate visual data with physical instability (vibrations or slip) experienced by the robot.
- VERN: A vision-based classifier that distinguishes between "pliable" vegetation (traversable grass) and "stiff" vegetation (impassable bushes or trees).
Enhanced Planner
The system extends the standard Dynamic Window Approach (DWA) with a novel objective function. This planner balances goal progress against costs derived from surface roughness, elevation changes, and vegetation density, rather than simply seeking the shortest geometric path.
Experimental Setup
The researchers evaluated AdVENTR using real-world data from two distinct platforms: a wheeled Clearpath Husky and a legged Boston Dynamics Spot.
- Training: Perception modules like TerraPN and VERN were pre-trained on datasets where robot-terrain interactions (IMU vibrations and odometry errors) served as self-supervised labels.
- Evaluation: The system was tested across six challenging real-world scenarios, including asphalt bordered by tall grass, uneven sandy pits, and rocky slopes.
Key Insights
- Priority-Based Perception: The system follows a hierarchy (Vegetation > Unevenness > Surface properties). For instance, if tall grass is detected, the vegetation module takes priority because it fundamentally redefines "traversable space."
- Proprioception as a Sensor: By performing Principal Component Analysis (PCA) on IMU data and joint forces, the robot can "feel" the difference between concrete and sand, allowing it to adapt even when visual cues are ambiguous.
- Dynamic Gait Adaptation: For legged robots like Spot, the system can trigger a switch to more stable gaits when high-risk surfaces, such as loose rocks, are detected.
Results
AdVENTR demonstrated significant improvements over standalone perception methods and commercial inbuilt planners.
- Success Rates: In a scenario involving rocky surfaces and tall grass, the Spot robot's inbuilt planner had a 0% success rate, as it treated the grass as an impassable wall. AdVENTR achieved an 80% success rate by correctly identifying the grass as traversable.
- Stability: Across all tests, AdVENTR consistently minimized "instability costs" (cumulative vibrations), resulting in paths that were physically safer for the hardware.
- Efficiency: AdVENTR generated significantly shorter trajectories than surface-only planners by recognizing that traversing pliable vegetation was more efficient than taking long detours on asphalt.
Conclusion and Future Work
AdVENTR provides a robust framework for complex outdoor navigation, yet some challenges remain. The system's reliance on RGB images for modules like VERN means performance can degrade in low-light conditions or during high-speed maneuvers causing motion blur. Additionally, low-hanging branches can occasionally be misinterpreted as ground-level obstacles in elevation maps.
Future work involves integrating Vision Language Models (VLMs) to provide higher-level semantic understanding. This could enhance the robot's ability to generalize to entirely new environments or varying weather conditions, further pushing the boundaries of autonomous outdoor exploration.
Summary updated