ACT-FLEX: A symbolic and generative AI integration architecture for generalisable and explainable robotic disassembly tasks

The increasing demand for personalised products poses new challenges for sustainable manufacturing, particularly for the large-scale disassembly required for circular economy initiatives. Traditional robotic systems lack the flexibility and generalisability needed to adapt across diverse product types and robot morphologies. In this work, we present ACT-FLEX, a cognitive architecture inspired by psychological models of human cognition. We used ACT-FLEX to show that pre-trained Large Language Models (LLMs) as well as Vision-Language Models (VLMs) can enable flexible, explainable and generalisable robotic disassembly across robot morphologies and tasks. Towards this, our architecture incorporates both symbolic reasoning, multimodal perception and morphology-aware action generation to support decision-making in dynamic environments. We validate ACT-FLEX across simulated and physical experiments involving multiple disassembly scenarios, product configurations and various robotic platforms including UR10, KUKA iiwa, Panda Franka and uArm Swift Pro. Results show transferability across robot platforms and robustness to product variations, with success rates of up to 80% in physical trials. This work demonstrates a step towards realising transparent, retaskable and sustainable robotic systems for Industry 5.0.