Neuro Symbolic Planning Algorithms For Complex Industrial Automation Sequences

Abstract

Planning for complex sequences of industrial automation requires a system capable of handling the uncertainties of the sensory environment, rigorous safety requirements, and traceability of the decisions made by an operator. Neural systems provide flexible perception capabilities, but do not guarantee the satisfaction of safety constraints, while traditional planners guarantee their correctness but cannot generalize over high-dimensional partially observed factory floors. We propose the Neuro-Symbolic Planning Algorithm (NSPA), which is a combined architecture of a Graph Neural Network (GNN) perception module, Hierarchical Task Network (HTN) planner, Z3 SMT solver and LLM-based re-planner. The SGL connects the neural and symbolic layers by performing the mapping between continuous latent space embeddings and discrete predicate spaces at any time. Our method is assessed on two industry benchmark environments - the realistic robot assembly environment, namely, AutoSim-v2, and the flexible manufacturing testbed, FlexMfg-Pro, where it achieves success rates of 95.3% and 94.1%, respectively, beating all the baselines significantly. Ablation study shows that each layer makes an effective contribution to the overall performance. Our approach works within the real-time requirement (an average planning delay of 0.43 s per sequence) and generates complete symbolic explanation traceable to safety-critical industry standard.