@PHDTHESIS{20.500.11850/700472,
	copyright = {In Copyright - Non-Commercial Use Permitted},
	year = {2024},
	type = {Doctoral Thesis},
	author = {Milojevic, Dejan},
	size = {242 p.},
	abstract = {The challenges of designing embodied intelligence systems arise from the necessity to integrate diverse hardware and software components without creating excessive complexity. The goal is to ensure safety and efficiency while minimizing resources such as costs, energy, computational requirements, and weight, considering the interactions between systems. This thesis presents a new approach for designing mobile robots tailored to specific tasks through automated selection of hardware and software. This approach emphasizes the interplay between perception systems and motion planning in decision-making. To evaluate and select perception systems for mobile robots, this work abstracts the performance of sensors and detection algorithms into False Negative Rate (FNR) and False Positive Rate (FPR) metrics. These metrics are evaluated considering various factors such as geometric relationships, object properties, sensor resolution, and environmental conditions, enabling a quantitative evaluation of different perception systems. To tailor the perception system for a specific task, it is crucial to determine what it needs to detect. This necessity introduces the concept of occupancy queries for sampling-based motion planners. Using these queries, we define the perception requirements based on the prior knowledge of object locations, dynamics, and shapes in the environment, demonstrating how to align the perception system with the demands of a given motion planner. After establishing a method to evaluate perception performance and identify perception requirements, this work connects these concepts by formulating the sensor selection and placement problem. This problem is solved as a weighted set cover problem using an Integer Linear Programming (ILP) optimization. The solution to this problem aims to select sensors and perception algorithms, and position the sensors to meet all perception requirements while simultaneously minimizing resources. This forms the basis for a co-design optimization that includes the robot body, motion planner, perception pipeline, and computing unit, utilizing the monotone theory of co-design. A case study on developing an Autonomous Vehicle (AV) for urban scenarios shows that tasks with higher complexity require more or equal resources. The extent of prior knowledge about object locations affects resource requirements, with numerous potential object positions leading to more elaborate designs. Additionally, the choice of motion planner influences the design: planners that need comprehensive environmental data require advanced and expensive sensors and perception algorithms. Moreover, enhanced task performance demands more resources, e.g., faster average speeds prompts the selection of vehicles with superior acceleration and motion planners that extract more information, thereby escalating resource demands. The study demonstrates that resource prioritization influences sensor choice: cameras are preferred for cost-effective and lightweight designs, while lidar sensors are chosen for better energy and computational efficiency. Lidars, with their superior perception performance and coverage, are essential for handling complex tasks.},
	keywords = {Co-design of Embodied Intelligence; mobile robots; Co-design; Sensor Selection and Placement; Perception; Motion Planning; Decision Making; Computer Vision; Autonomy; Autonomous Vehicles; Set cover problem; Optimization; Benchmarking},
	language = {en},
	address = {Zurich},
	publisher = {ETH Zurich},
	DOI = {10.3929/ethz-b-000700472},
	title = {Co-Design of Mobile Robots: Integrating Perception Systems and Motion Planning for Task-Specific Optimization},
	school = {ETH Zurich}
}