RED scheduler keeps robot AI on time as tasks shift in real-time

Robots operating in dynamic environments face a tough challenge: the computation graph itself can change at runtime as new tasks appear, precedence relations shift, or sensor inputs become unpredictable. This wreaks havoc on traditional real-time schedulers, especially when multi-task deep neural network inference must squeeze into tight resource and timing budgets. A new paper from Zexin Li and colleagues introduces RED (Robotic Environmental Dynamics), a deadline-aware scheduler designed specifically for this scenario. Its core innovation is assigning intermediate sub-deadlines to portions of the DAG-based workload, allowing it to gracefully accommodate evolving computation graphs and asynchronous inference triggered by environmental changes. RED also explicitly exploits the structure of MIMONet (multi-input multi-output neural networks), which use weight sharing to reduce memory pressure — a common design in multi-tasking robots. The framework includes a workload refinement and graph-reconstruction step that aligns the MIMONet structure with schedulability requirements, improving both efficiency and compatibility. The team implemented RED on NVIDIA Jetson platforms and an Apple M-series MacBook, testing it with realistic navigation workloads that include object detection, path planning, and control inference. Results consistently beat existing methods across throughput, deadline satisfaction, robustness to interference, adaptability, and runtime overhead — suggesting a practical path for next-gen autonomous robots that need to react intelligently in unpredictable environments.