Robustness analysis of genetic programming controllers for unmanned aerial vehicles

Barlow, Gregory J.; Oh, Choong K. · 2006 · Crossref

DOI: 10.1145/1143997.1144023

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical challenge of ensuring the safety and reliability of evolutionary robotics controllers when transferring them from simulation to real-world unmanned aerial vehicles (UAVs). Because poorly performing controllers can cause crashes in physically unstable systems like fixed-wing UAVs, the authors argue that evolved controllers must be robust to environmental noise and sensor inaccuracies. The study focuses on selecting the best navigation controllers from a large population generated by multi-objective genetic programming (GP) and verifying their performance under various off-design conditions prior to real flight tests. The researchers evolved autonomous navigation controllers for fixed-wing UAVs tasked with locating, tracking, and orbiting radar sites. Using an implementation of NSGA-II for GP, they performed 50 evolutionary runs with 500 individuals each, generating 25,000 controllers. The evolution utilized environmental incremental evolution, progressively increasing task difficulty across four radar types: continuously emitting stationary, continuously emitting mobile, intermittently emitting stationary, and intermittently emitting mobile radars. The simulation modeled realistic constraints, including a constant speed of 80 knots, specific autopilot update rates, and sensor noise (±10° angle of arrival and ±6dB amplitude). To select candidates for robustness testing, the authors applied a multi-stage filtering process based on four fitness functions measuring flight time, circling distance, flight efficiency, and stability. This reduced the pool to 10 top-performing controllers, which were then compared against two hand-designed controllers: a domain-specific heuristic controller and a proportional-derivative (PD) controller. The selected controllers underwent rigorous robustness tests involving increased sensor noise and introduced state noise, including variations in airspeed, heading error, and wind effects. Performance was evaluated using four test functions derived from the fitness metrics, normalized against baseline values for a minimally successful controller. The results demonstrated that the best evolved GP controllers significantly outperformed the hand-designed controllers. Specifically, the evolved controllers maintained stable and efficient flight paths, successfully tracking radars even under high levels of noise and varying environmental conditions. The robustness analysis confirmed that controllers evolved in noisy simulations could handle a wide range of sensor and state errors, provided the real-world noise fell within the tested ranges. The significance of this work lies in providing a systematic methodology for validating evolved controllers for high-risk applications. By demonstrating that GP-evolved controllers can be robust to noise and outperform traditional hand-designed solutions, the paper supports the feasibility of transferring evolutionary robotics controllers to real UAVs. The findings suggest that evolving controllers in simulations with diverse noise conditions is an effective strategy for ensuring safe and reliable real-world performance, thereby addressing a major bottleneck in the deployment of autonomous evolutionary robotics systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• motorcyclist skill