Under-Canopy Navigation using Aerial Lidar Maps

TL;DR

This work tackles under-canopy ground navigation by leveraging above-canopy aerial lidar data to build a probabilistic 3D occupancy map that accounts for pose uncertainty. It introduces Monte Carlo-based uncertainty propagation to fuse aerial measurements into a 3D map, derives a 2D ground-obstruction score, and integrates this into two cost functions for a D* Lite global planner, enabling efficient and replannable paths. The approach is validated through extensive simulations with ablation studies and real-world experiments in dense forests, showing improved map accuracy (measured by $p(m_i|oldsymbol{\Z})$) and shorter, safer trajectories using a log-reachability-based planning objective. Overall, the method demonstrates that combining uncertainty-aware aerial priors with risk-aware planning substantially aids robust ground navigation in complex, canopy-covered environments.

Abstract

Autonomous navigation in unstructured natural environments poses a significant challenge. In goal navigation tasks without prior information, the limited look-ahead of onboard sensors utilised by robots compromises path efficiency. We propose a novel approach that leverages an above-the-canopy aerial map for improved ground robot navigation. Our system utilises aerial lidar scans to create a 3D probabilistic occupancy map, uniquely incorporating the uncertainty in the aerial vehicle's trajectory for improved accuracy. Novel path planning cost functions are introduced, combining path length with obstruction risk estimated from the probabilistic map. The D-Star Lite algorithm then calculates an optimal (minimum-cost) path to the goal. This system also allows for dynamic replanning upon encountering unforeseen obstacles on the ground. Extensive experiments and ablation studies in simulated and real forests demonstrate the effectiveness of our system.

Figures (8)

• Figure 1: Sensors mounted on a ground robot can suffer from limited range and occlusions. This local view often precludes efficient long-distance navigation in complex scenarios, leading to dead-ends or unsafe paths (right figure: yellow arrows). Aerial lidar data (top left figure) acquired prior to operation can provide meaningful information for global planning (bottom left and right figures: magenta paths) allowing efficient navigation.
• Figure 2: Overview of our proposed navigation system. First, aerial lidar data are fused with uncertain sensor poses to generate a 3D probabilistic occupancy map. Next, the 3D map is further processed to generate a 2D ground obstruction map. Finally, obstruction scores are passed through a cost function to the D* Lite planner to produce global paths that guide the ground robot during operation.
• Figure 3: We reason on free, occupied and uncertain voxels by integrating lidar measurements in an occupancy grid map. Later the occupancy probabilities are fused in our scoring method to estimate the final obstruction map.
• Figure 4: (a): Example of synthetic forest created in Gazebo where we simulated lidar scans captured above the canopy. (b) to (d): Ground-truth trajectories (A, B and C) of the airborne sensor and the corresponding registered scans for the datasets acquired in the synthetic environments Forest I, II and III, respectively. After adding perturbations to sensor poses, a perturbed point cloud of each dataset was produced to be processed by our approach.
• Figure 5: Comparison of our UA-occupancy against the standard occupancy map for different perturbation values added to sensor poses. (a) and (b) show the average KLD values in Forest I(trajectory A) for perturbations in position only and orientation only, respectively. KLD values are depicted per voxel class on the ground-truth dataset. (c) and (d) illustrate the KLD values for all voxel classes in Forest II(2 trajectories) and Forest III(2 trajectories), respectively, for combined perturbation levels (position and orientation) where $\sigma=(0.015 \text{m}, 0.015 \text{m}, 0.015 \text{m}, 0.11^{\circ} , 0.11^{\circ}, 0.11^{\circ})$. In the figures, noise refers to perturbations added to sensor poses and error bars show $\pm$ std. deviation.