AutoPilot: Compute Design Automation for Autonomous Drones
TimeWednesday, July 13th6pm - 7pm PDT
LocationLevel 2 Lobby
DescriptionAutonomous drones are complex cyber-physical systems, and designing onboard compute for these platforms is challenging due to a lack of systematic design methodologies. Iterating over component choices (sensors, compute, algorithms, physical dimensions) results in a combinatorial explosion in the number of possible designs, making ad-hoc selections intractable. To address this problem, we introduce AutoPilot, a comprehensive ML-based automated cyber-physical co-design framework for compute design automation in autonomous drones. AutoPilot-generated designs increase mission QoS metrics by up to 2.3x, 1.57x, and 1.33x on average for nano, micro, and mini-UAVs, respectively, compared to Xavier NX, Jetson TX2, and dedicated hardware accelerators.