CAMTEC: Context-aware Adaptive Inference for Machine Task Execution in Complex Environments

The rise of smart factories is transforming manufacturing and service delivery using robots and unmanned ground vehicles (UGVs) for advanced tasks like autonomous navigation. While over-the-top (OTT) solutions reduce local compute and energy demands by offloading machine learning (ML) tasks to remote servers, they lack task-aware optimization. This limitation is critical for complex applications like autonomous navigation, where performance depends on dynamic device and network conditions. To bridge this gap, we propose Context-aware Adaptive inference for Machine-Task Execution in Complex environments (CAMTEC). It incorporates performance metrics from compute, energy, network, and ML to generate the context of machine task performance (e.g., UGV navigation). CAMTEC utilizes a deep reinforcement learning (DRL) agent to learn from this context information to determine the optimal inferencing mode (local, remote, or split) and vehicle speed to ensure optimal performance of an off-the-shelf object detection-based navigation algorithm. Extensive real-world testbed evaluations demonstrate CAMTEC’s efficacy in achieving collision-free autonomous navigation. CAMTEC adheres to a computationally optimal allocation strategy throughout the vehicle’s trajectory, ensuring that the actual-to-target speed variation remains below 20% even under battery-constrained conditions. Furthermore, even when network bandwidth is severely limited, CAMTEC ensures that the UGV converges to the optimal path while keeping the actual-to-target speed variation below 30%.