The rise of autonomous machines, including autonomous vehicles, robots, and drones, is revolutionizing industries and daily life. These systems, especially autonomous vehicles, rely on powerful computing platforms to make split-second decisions, such as detecting obstacles, adjusting speed, and coordinating movements. However, making smart decisions is not enough; these machines must also act with precise timing to ensure safety and reliability. For instance, an autonomous vehicle approaching an intersection must identify a pedestrian immediately and apply the brakes without delay - even a few milliseconds of lag could mean the difference between a safe stop and a collision. This demand for strict timing guarantees makes real-time computing a critical challenge in autonomous driving systems.
To ensure safe and predictable operation, autonomous driving systems must pass real-time safety certifications, which rely on two key analyses. Worst-Case Execution Time (WCET) measures the longest possible time a task might take to complete, ensuring the system accounts for delays under extreme conditions. Schedulability Analysis ensures all tasks meet their deadlines, even under heavy workloads. However, today’s computing technologies struggle to provide reliable timing guarantees, particularly when using artificial intelligence (AI) accelerators. These specialized chips, such as GPUs and ASIC AI accelerators, significantly speed up complex tasks like image recognition and decision-making but introduce new uncertainties. Current systems face three main challenges: modern SoCs lack built-in support for real-time scheduling, AI accelerators create unpredictable execution times, and existing methods fail to estimate AI task timing accurately due to the variable nature of deep neural network (DNN) computations.
Our research proposed in this project addresses these challenges by integrating AI acceleration with real-time computing principles through three key innovations. First, we are developing an accelerator-enhanced SoC designed specifically for real-time scheduling, ensuring that critical DNN tasks complete on time while optimizing overall efficiency. Second, we are building a real-time scheduling framework that understands and predicts DNN execution behavior, allowing operating systems to make smarter scheduling decisions and prevent unpredictable delays. Third, we are introducing an outlier management strategy to improve WCET estimation for DNN tasks, ensuring real-time safety standards are met even in unpredictable conditions.
Solving these timing challenges will be a major breakthrough, making AI-powered autonomous driving systems more reliable and safe. Furthermore, our innovations will not only enhance autonomous vehicles but also improve the performance of robots, drones, and industrial automation systems that rely on precise real-time execution. By bridging the gap between AI acceleration and real-time computing, our research lays the foundation for the next generation of safe, predictable, and efficient autonomous technologies that can seamlessly integrate into everyday life.
Last modified 25 February 2025