Hardware-Adaptive Learning-Assisted Predictive Control Architecture for Real-Time Trajectory Planning

Abstract

The high level of real-time trajectory control in embedded autonomous systems requires very strong timing specifications, resilience to model requirement, and economic use of severity computer facilities. Although the model predictive control (MPC) enables optimal control performance with explicit constraint management, it has a high computational cost, which makes it extremely difficult to implement on a limited resources embedded system using a real-time application. Assisted control methods based on learning have also been investigated to enhance the accuracy and flexibility of prediction, but they have the disadvantage of non-deterministic implementation and unpredictable latency, which makes them less suitable when application in safety-critical systems. This paper introduces a hardware-adaptive learning-assisted predictive control architecture that has been developed to facilitate deterministic real-time trajectory planning by closely coordinating the functions of learning-enhanced prediction with reconfigurable hardware execution. The suggested methodology will be based on a hardware-software co-design model, whereby, time critical MPC computations are assigned to reconfigurated hardware accelerators, with a lightweight learning component closing model parameters and prediction accuracy during runtime. A hardware adaptation scheme is a dynamically reconfigurable system that has computational resources which are reallocated based on workload and timing requirements, guaranteed worst-case execution time whilst maximizing performance and energy efficiency. The suggested architecture is realized and tested on a specific embedded platform of an FPGA platform with a real-time situation in trajectory planning. Expertise findings indicate significant control latency and energy usage cutbacks together with better trajectory following precision over software solitary and fixed hardware reference designs. These findings approve the fact that the suggested hardware-adaptive learning-assisted architecture is successful in coping with the complexity of real-time embedded trajectory planning and is appropriate to the safety-critical autonomous and cyber-physical systems.