Enhancing Imitation Learning for Robotic Assembly with 6D Pose Estimation and Reinforcement Learning Fine-Tuning

Robotic manipulation constitutes a fundamental domain in robotics research, with robotic assembly-an essential subset-rapidly evolving and presenting considerable challenges. Unlike other manipulation tasks, robotic assembly demands high physical complexity and reliability, making full automation especially difficult. To address these issues and enhance the performance of robotic assembly tasks, this paper proposes a novel imitation learning training and inference framework that integrates reinforcement learning (RL) fine-tuning and six degrees-of-freedom (6D) object pose estimation on top of a behavior cloning backbone. Additionally, a privileged learning paradigm is adopted to alleviate hardware demands during training. Our method is applied to an assembly task involving long horizon and multi-stage precise grasping and part insertion. In simulation, we train and evaluate the full assembly pipeline. Our method substantially outperforms baseline methods, improving the success rate of the one-leg assembly task from 54% to 70%-an absolute increase of 16%.