This thesis presents an intellectualisation solution for an automated flexible sheet metal forming process, which produces arbitrarily curved sheet metal components through incremental local deformation. For the first time, an intelligent flexible sheet metal forming system is established to provide an end-to-end tool path planning framework, with state-of-the-art theory-guided/data-driven artificial intelligence (machine learning) technologies embedded for tool path fast prediction and system self-improvement. The system is composed of seven parallel modules covering data storage, raw data pre-processing, tool path planning, process computation and validation, post-processing and feedback mechanism, of which the tool path planning accuracy and efficiency is ensured through a dual modular redundancy-akin approach. The path planning problem is investigated through three stages: target sheet deformed by 1) a single punch, 2) multiple punches, and 3) unlimited punches. At the first stage, a classic convolutional neural network (CNN) is developed for process parameter prediction of bending processes. A novel theory-guided regularisation method, utilising Swift’s law as guidance, for deep neural network (DNN) training is proposed, which has been verified to precede conventional data-driven method in severely scarce data condition. The theory-guided DNN exhibits a more robust generalisation capability and learning consistency than the data-driven DNN in experiments with different training data structures, workpiece materials and sheet metal forming applications. At the second stage, a consecutive rubber-tool forming process is developed to incrementally deform the sheet metal, whose tool path is predicted through a recursive tool path prediction framework. The framework adopts a deep learning model to learn the topological relationship between the current and target workpiece shape. Three series of state-of-the-art deep learning models, namely single feature extractor, cascaded networks and long short-term memory (LSTM) models are investigated. With over 6000 samples, the CNN LSTMs have been verified to be the most superior in tool path planning owing to their ability in learning both spatial and temporal features, while the other two models exhibit inferior performances when path planning complexity increases. At the third stage, deep reinforcement learning (DRL) is exploited to learn the optimal tool path, without expertise in prior, for arbitrarily curved sheet metal. In this context, a generalisable tool path planning strategy is proposed to address the commonly-known low generalisation issue of pure DRL approach. The strategy factorises workpiece into segments and optimises tool path through dynamic programming, exploiting both DRL and deep supervised learning (DSL) technologies. The strategy manifests self-learning characteristics during training, and it has been verified applicable to tool path planning of arbitrary sheet metal parts through a case study. With the investigation results and dedicated tool path planning strategy in this thesis, the intelligent flexible sheet metal forming system can be deployed for fast and accurate tool path generation for prototyping during the sheet metal production cycle.