Romanus: Robust Task Offloading in Modular Multi-Sensor Autonomous Driving Systems

Such a multi-sensor approach leads to the generation of an enormous volume of high-dimensional data that requires tremendous Due to the high performance and safety requirements of self-driving resources for real-time processing, further adding to the power applications, the complexity of modern autonomous driving systems demands of the entire system. Addressing this, a heterogeneous (ADS) has been growing, instigating the need for more sophisticated collection of hardware components, as in Application-Specific Integrated hardware which could add to the energy footprint of Circuits (ASICs) and GPUs, are commonly integrated onto the ADS platform. Addressing this, edge computing is poised to encompass ADS platforms to balance performance demands and power efficiency self-driving applications, enabling the compute-intensive [13]. Still, hardware advancements are met with growing autonomy-related tasks to be offloaded for processing at computecapable algorithmic complexity and the requirement for supporting new edge servers. Nonetheless, the intricate hardware architecture features, leading the power footprint to remain relatively high. For of ADS platforms, in addition to the stringent robustness instance, if we compare two generations of ADS platforms: the demands, set forth complications for task offloading which are Nvidia Drive PX2, which was used by Tesla and Audi Q7 for their unique to autonomous driving. Hence, we present ROMANUS, a autopilot programs [1, 2], against its successor, the Nvidia Drive methodology for robust and efficient task offloading for modular AGX Orin [4], we find that performance efficiency aside, the baseline ADS platforms with multi-sensor processing pipelines. Our methodology power demands increased from 250 W to 800 W, which in entails two phases: (i) the introduction of efficient offloading theory can have adverse effects on both the thermal comfort of the points along the execution path of the involved deep learning models, passengers and the vehicle's driving range [13].