Almeida, A, Rueda, DA, Haque, NI et al. (2026). RAPTOR: Adversarially Robust Path Planning for LiDAR-Based Autonomous Robotic Vehicles
. 3128-3133. 10.1109/COMPSAC69091.2026.00464
Almeida, A, Rueda, DA, Haque, NI et al. (2026). RAPTOR: Adversarially Robust Path Planning for LiDAR-Based Autonomous Robotic Vehicles
. 3128-3133. 10.1109/COMPSAC69091.2026.00464
Autonomous robotic vehicles (ARVs) are increasingly deployed in safety-critical environments where reliable path planning is essential. Although industry systems commonly use multi-sensor fusion with camera-based perception, these approaches raise privacy and ethical concerns due to potential capture of personally identifiable information, motivating privacy-preserving navigation using non-imaging sensors such as LiDAR. Classical graph-based planners (e.g., A*) provide optimal and deterministic paths under clean sensing conditions; however, they rely on accurate sensor-derived representations and remain vulnerable to LiDAR spoofing and false data injection, particularly in dynamic environments requiring frequent replanning. ML-based path planners provide faster, adaptive decision making but introduce additional vulnerabilities to adversarial attacks on LiDAR inputs. To address these challenges, we propose RAPTOR, a security framework for LiDAR-based ARV path planning that models dynamic environments and two realistic adversary classes. RAPTOR combines adversarial training, temporal modeling through long short-term memory planners, and runtime ensemble disagreement detection with a verifiable A∗ fallback. Extensive evaluation in dynamic environments shows that RAPTOR improves robustness, reduces collision rates, and maintains real-time performance for secure, privacy-preserving LiDAR-based navigation under diverse adversarial conditions.
