Ashiqur Rahaman M.

Safe execution in perception-based autonomous systems requires integrating AI perception and control tasks. We propose a novel methodology to co-optimize these on a real-time CPU-GPU platform. Moving beyond isolated designs, we link overall performance to both perception inference accuracy and iterative control execution. We validate this design-space exploration methodology on an autonomous driving use case using a novel simulation setup.

Future SDVs rely on zonal architectures linking diverse legacy protocols (CAN, FlexRay) via a Time-Sensitive Ethernet backbone. Current inter-zone routing is ad-hoc, wasting bandwidth. We propose an optimal SMT formulation that multiplexes periodic zonal messages into minimal Ethernet frames using (m,k)-firmness relaxations, while efficiently routing them. Z3 solver evaluations demonstrate a 30% gain in frame utilization for robust SDVs.