Perception accuracy in physical AI systems for autonomous driving is increasingly limited by slow, hardware-bound iteration across the sensor-to-silicon pipeline, rather than by model capacity. In production programs, most of the development effort is spent outside neural network training - on tuning, optimization, data regeneration, and hardware-bound validation - often revealing accuracy regressions months too late. In this talk, Arm, in collaboration with autonomous mobility start-up covolv.ai, demonstrate how to achieve faster perception convergence, earlier verification, and scalable ODD iteration by integrating end-to-end virtualized compute directly into training and validation loops - well before silicon availability.

• What are the common mistakes in image pre-processing that impact engineering time, compute and model accuracy
• How to achieve faster perception model convergence
• What are the benefits of virtualised verification at scale