Multi-Objective Generation of Pareto-Optimal Perception Architectures for Autonomous Robotic Systems

Abstract

Designing perception systems for autonomous robots and vehicles requires balancing sensor performance against cost, complexity, and integration constraints. This thesis introduces GO4R (Generation and Optimization of Perception System Architectures for Robotics), a multi-objective framework that jointly designs sensor selection, placement, against volumetric, entropy-based utility metric H (-) and monetary cost M ($). Perception Entropy H is formalized as a volumetric measure of uncertainty across a voxelized regions of interest (ROI), which naturally rewards coverage, overlap, and redundancy required for robust sensor fusion and calibration.

NSGA-II is implemented with custom mixed-variable operators to specifically handle both continuous (e.g. sensor poses) and discrete (e.g. sensor type/count) decision variables found in this problem. Two case studies, long-range outdoor navigation on a Clearpath Jackal and short-range indoor navigation on ANYmal-C, demonstrate the framework’s ability to generate Pareto-optimal sensor architectures under vastly different ROI definitions and operating conditions. In the Jackal study, GO4R converges to a population of 11 novel Pareto-optimal designs, and revealing sensitivity to voxel size and importance weighting. In the ANYmal-C study, the compact, uniformly weighted ROI yields a flatter Pareto front with 25 Pareto-optimal designs, and underscores how intrinsic sensor parameters (e.g. angular resolution, and Field of View) dominate design trade-offs when baseline coverage is already high.

Key architectural decisions are analyzed, quantified by their impact on Pareto front shape, and ordered according to the GO4R method to successively reduce uncertainty. The resulting guidelines provide practitioners with a rigorous, reusable process for tailoring perception systems to task-specific requirements. Finally, GO4R provides a publicly available NVIDIA Isaac Sim extension to aid practitioners in following the GO4R method, no matter their Autonomy application. Future work will extend GO4R to dynamic environments, improve fidelity of generated designs, and incorporate additional cost metrics such as computational load and maintainability.