Bringing Anytime Perception to Real Hardware: An Embedded Deployment of the Autoware Stack with Dynamic Resolution Scaling

Deploying deep neural networks for perception on autonomous vehicles forces a compromise between how accurately the system perceives and how quickly it responds. This compromise is especially binding on embedded compute platforms, where limited processing power means a high-accuracy detector may fail to finish within the control loop's timing budget, leaving the vehicle to act on outdated information. Anytime perception offers a way to manage this by adjusting inference cost at runtime, but its benefits have so far been shown mainly in simulation, with little evidence from physical deployment.

This thesis provides that evidence. We take MURAL — a multi-resolution anytime LiDAR detector previously integrated into the Autoware stack and evaluated in the AWSIM simulator — and deploy it on a physical mid-size rover, running the full sensing-to-actuation pipeline on a single NVIDIA Jetson AGX Orin. Reaching a working deployment required substantial adaptation of a stack originally built for full-scale vehicles in simulation, from retargeting the vehicle model to rover scale to bringing the entire pipeline on-board a single embedded device.

By carrying the complete stack onto real hardware, this work makes it possible to evaluate anytime perception under the conditions it was designed for: a full autonomous-driving pipeline running on an edge device in the physical world. We assess, through end-to-end physical experiments, whether dynamically scaling detection resolution delivers a real performance benefit on embedded hardware — providing, to our knowledge, the first true evaluation of anytime perception for edge-deployed autonomous driving.