Decentralised task planning and motion coordination for scalable multi-robot collaborative manufacturing

This paper introduces a decentralised task planning and motion coordination (TPMC) framework for scalable multi-robot collaborative manufacturing (MRCMfg). Built upon cognitive digital twins, the framework addresses challenges in dynamic system configurations, such as real-time adjustments to the number and position of active robots. The manufacturing process is divided into four phases: cognition, configuration, planning, and coordination. A large language model (LLM) analyses and infers data from cognitive digital twin simulations for state cognition and configuration optimisation, ensuring scalability by dynamically adjusting robot layouts. Multi-agent deep reinforcement learning (MADRL) is employed for TPMC, which relies solely on local observations and incorporates a Transformer, along with a Soft Mixture-of-Experts (MoE) module, to update the policy network. This results in a decentralised decision-making policy with superior scalability and generalisation capabilities. A novel decentralised Soft-Soft Actor-Critic (SSAC) algorithm is developed, integrating a Soft MoE into the policy network, enhancing policy generalisation. The method is trained and tested on a scalable MRCMfg system in cognitive digital twins, specifically for the disassembly of electric vehicle batteries. Experiments demonstrate the approach's effectiveness, scalability, and efficiency in handling dynamic system configurations, as well as achieving efficient TPMC. The paper concludes by highlighting the framework's ability to maintain continuous operation in dynamic environments. The paper also suggests future work on enhancing the coupling mechanisms between task and motion planning (TAMP), integrating stochastic task-duration modelling, and incorporating physical priors into training.