Task attention-based multimodal fusion and curriculum residual learning for context generalization in robotic assembly
摘要
In the domain of flexible manufacturing, Deep Reinforcement Learning (DRL) has emerged as a pivotal technology for robotic assembly tasks. Despite advancements in sample efficiency and interaction safety through residual reinforcement learning with initial policies, challenges persist in achieving context generalization amidst stochastic systems characterized by large random errors and variable backgrounds. Addressing these challenges, this study introduces a novel framework that integrates task attention-based multimodal fusion with an adaptive error curriculum within a residual reinforcement learning paradigm. Our approach commences with the formulation of a task attention-based multimodal policy that synergizes task-centric visual, relative pose, and tactile data into a compact, end-to-end model. This model is explicitly designed to enhance context generalization by improving observability, thereby ensuring robustness against stochastic errors and variable backgrounds. The second facet of our framework, curriculum residual learning, introduces an adaptive error curriculum that intelligently modulates the guidance and constraints of a model-based feedback controller. This progression from perfect to significantly imperfect initial policies incrementally enhances policy robustness and learning process stability. Empirical validation demonstrates the capability of our method to efficiently acquire a high-precision policy for assembly tasks with clearances as tight as 0.1 mm and error margins up to 20 mm within a 3.5-hour training window-a feat challenging for existing RL-based methods. The results indicate a substantial reduction in average completion time by 75