Towards Modernization of Long-Range Image-Space Planning for Off-Road Navigation

Abstract

This thesis revisits long-range, image-space planning for off-road navigation and modernizes the classical first-person view (FPV) paradigm by building upon recent advances in perception. It introduces a lightweight depth calibration scheme, analytic configuration-space (C-space) transforms, interpretable frontier selection, and a pixel-space A* planner with validated heuristic soundness. Concretely, we (i) make monocular depth metrically usable at test time via an affine, log-domain calibration with sparse LiDAR; (ii) derive and implement a depth-aware FPV C-space inflation that projects vehicle width/length analytically and realizes it with separable row/column sliding-maximum filters augmented by per-pixel depth consistency checks; (iii) propose transparent, angular-sector frontiering that reasons jointly about traversability cost and minimal lethal depth, alongside goal-aware revalidation; and (iv) preserve A* admissibility/consistency in image space through a simple cost renormalization that avoids silent suboptimality in low-cost free space.

We evaluate the resulting, modular sub-stack in the high-fidelity Falcon simulator [2] under a shared ROS graph. Using a common perception front-end and planning back-end, we compare three frontiering strategies -- (1) a LAGR-style row-wise baseline, (2) an LRN-inspired openness heuristic adapted to operate with explicit depth and cost, and (3) an Angular Cost & Depth (ACD) variant that couples average cost with a minimum lethal-depth constraint. Across diverse courses (e.g., farm, desert, mixed terrain), the calibrated monocular depth reduces error versus raw monocular predictions, and both LRN-inspired and ACD frontiering tend to outperform the purely row-wise baseline on longer traverses. We view these as encouraging indications rather than definitive claims: all results are in simulation, with performance subject to perception quality, calibration coverage, and environment diversity.

Scope and limitations are explicit. The work was conducted over a short project window (May 2025--Oct 2025) and prioritized stabilizing the proposed sub-stack in conjunction with a core FieldAI [1] stack through the high-fidelity Falcon simulator [2] and ROS1 into a reliable, end-to-end testing framework. No real-world deployments were performed. Nevertheless, the design is intentionally modular and auditable to make it relatively straightforward to integrate the sub-stack into existing off-road autonomy stacks lacking explicit long-range planning. Such integration might enable better autonomy by offering a practical bridge between classical image-space efficiency and metric-world robustness.

[1] FieldAI: https://www.fieldai.com
[2] Falcon: https://www.duality.ai