Data-Driven Disassembly Line Allocation via RoBERTa Fine-Tuning and Reinforcement Learning

In modern manufacturing, efficient disassembly line allocation is critical for productivity and recycling value but is often hindered by manual bottlenecks. To address this, this work proposes RoBERTa-based disassembly allocation system (RoDAS), a unified framework synergizing RoBERTa-based semantic perception and reinforcement learning (RL) control. By leveraging a pretrained language model to generate task-specific datasets, RoBERTa is fine-tuned to process complexity-augmented prompts, integrating semantic features and product priors to generate precise initial policies. This encoder integrates graph-based knowledge retrieval and vector databases to construct multisource prompt, enabling the generation of a precise, knowledge-driven initial policy. Subsequently, an RL agent utilizes this strategy as a warm-start mechanism to simulate the disassembly process and dynamically calibrate decisions under real-time constraints. Extensive experiments validate that RoDAS achieves a Pareto-optimal balance, significantly reducing completion time while maintaining recycling value compared to leading state-of-the-art methods, including DeepSeek and GPT-4o.