Perception Stack Design and Fault-Aware Sensing for the EcoCAR Competition

Abstract

A major challenge in the development of autonomous driving systems is ensuring that the perception layer reliably interprets and manages sensor data under real-world operating conditions. In competition and research contexts where sensors are integrated onto production vehicles, this challenge is compounded by the vehicle's existing architecture where sensors expose processed detections, rather than raw data, over proprietary CAN buses, shifting the design focus toward robust data ingestion, active health monitoring, and failure management. This thesis presents the design and implementation of a perception stack for the University of Waterloo Alternative Fuels Team competition vehicle, a 2023 Cadillac LYRIQ, developed as part of the EcoCAR Electric Vehicle Challenge. The system operates on the dSPACE AUTERA platform using RTMaps middleware and implements a modular, per-sensor pipeline comprising CAN burst detection and processing, loss-of-communication monitoring, frame decoding, sensor health checking, and data processing. A fail-safe module was designed and implemented as a custom RTMaps Python bridge component using a fault-streak counting strategy that distinguishes sustained fault conditions from transient anomalies before issuing a disable command to the planning module so that features dependent on a compromised sensor are deactivated to ensure safety. The system was validated through the XIL testing framework, progressing from synthetic CAN frame generation and baseline vehicle evaluation data replay through hardware-in-the-loop testing on the AUTERA, demonstrating that a modular, fault-aware perception pipeline can be designed and implemented on a production vehicle platform under competition constraints.