Data-Efficient Learning-Assisted Predictive Control for Real-Time Trajectory Planning Under Dynamic Constraints

Abstract

Planning real-time trajectories of autonomous systems in real-world, recognised as dynamic and uncertain conditions is a highly challenging task because of dynamical nonlinearities, safety requirements, real-world computational constraints, and data intensity of learning-based solutions. Although model predictive control (MPC) offers methodical management of constraints and stability assurances; model discrepancy and complexity issues in real-time implementations usually impair its performance. On the other hand, the control techniques that are based on learning have flexibility to uncertainties but are often inefficient with data, and do not provide formal certainties of safety. In its attempt to overcome these drawbacks, this paper presents a Data-Efficient Learning-Assisted Predictive Control (DEL-APC) framework that is based on the synergistic implementation of lightweight learning mechanisms and constraint-conscious predictive control to plan the safe and real-time trajectory in dynamically constrained environments. The framework proposed incorporates a receding-horizon MPC formulation with a few-shot learning-based dynamics correction geared towards achieving rapid online adaptation with minimal data at its disposal as well as model interpretability. In order to guarantee sound constraint satisfaction and closed loop stability when learning uncertainty is there, a learning-conscious constraint tightening policy is added to the predictive optimization procedure. This construction ensures that it is recursively feasible and can be operated safely without being adversely affected on its computational efficiency. The efficiency of the suggested DEL-APC is confirmed by a numerical amount of simulation investigation of autonomous ground vehicle and quadrotor systems exposed to disturbances, time-varying, and modelling uncertainties. Experimental findings support the claim that DEL-APC enables a maximum decrease of trajectory tracking error of up to 45 percent with less than 60 percent of training performance than conventional learning-based controllers, and is not lenient to state and input constraints. Moreover, the suggested approach is guaranteed to result in real-time execution and control cycles time less than 5 ms, which is much superior to conventional MPC and deep reinforcement learning benchmarks. These findings indicate the promise of DEL-APC as a viable and dependable control system in safety-critical systems that call autonomous systems to use complex environments that are data-constrained.