Pipeline-Oriented Learning-Assisted Predictive Control for Real-Time Trajectory Planning in Large-Scale Autonomous Systems

Abstract

Robotic swarms and distributed cyber-physical infrastructures are large-scale autonomous systems that need real-time trajectory planning with severe limitations on latency, communication, and computational resources. Predictive control assisted by learning has proven to be an encouraging solution to improving flexibility and predictive accuracy, but the majority of the currently available solutions use highly integrated architectures, which cannot be scaled and pose difficulties with implementation in distributed systems. This paper suggests a pipeline-oriented predictive controller based learning architecture of autonomous systems in large scale and real-time trajectory optimization. The suggested methodology breaks the control workflow down into pipeline stages that are modules, which include state acquisition, learning aided short horizon prediction, trusting information filtering, predictive control optimization, and actuation feedback. This decomposition allows parallel and distributed execution among heterogeneous computing resources and such decomposition remains real-time feasible. Learning elements are used as predictor aids in the state estimation and predicting system activity, but are not used in place of an underlying model-based control framework. Besides this, a trust-aware, mechanism is added to reduce the impact of unreliable, delays or corrupted information in distributed communication environments. The broadly spread simulation based tests in dynamic conditions and in degraded communicating settings have verified that the proposed pipeline-based approach is better in the regards of reduced execution latency, scalability and robustness in the control of trajectory when compared to the traditional monolithic approaches that use learning and robust predictive control. These findings suggest that pipeline automation will provide a viable and scalable entry point that will support the implementation of learning-assisted predictive control in large-scale autonomous systems with real-time performance.