Strategy Optimization and Imitation Learning for Bimanual Robots Based on Hierarchical Encoder-Decoder (HED)

Fine manipulation tasks, such as precise grasping, insertion, and assembly, pose significant challenges for robots due to the need for accurate contact force coordination, real-time visual feedback, and dynamic trajectory adjustments. Existing solutions typically rely on high-end robots and sophisticated sensors, which are expensive and complex to configure. To address this, we propose a strategy optimization method based on a Hierarchical Encoder-Decoder (HED) model. HED improves the success rate and execution stability in complex environments through multi-layer feature extraction and dynamic trajectory generation. The encoder captures low-level spatial details, mid-level object relationships, and high-level task dynamics, enhancing adaptive strategy generation. The decoder generates smooth and precise trajectories using action chunking and temporal ensemble mechanisms, effectively reducing compounding errors. In simulated tests, HED achieved a success rate of 85% to 92% in complex manipulation tasks, significantly outperforming traditional methods. These results demonstrate the potential of HED for enabling low-cost robots to perform high-precision fine manipulation tasks.