Action chunking transformer-based imitation learning with dynamic programming optimization for low-cost robotic assembly tasks

Abstract

Nowadays, many studies focus on using AI tools to improve the performance of manipulators in fields such as assembly. One popular approach is the Action Chunking Transformer (ACT), which enables robots to execute assembly tasks more efficiently. However, this method still faces challenges in achieving reliable assembly performance, particularly for low-cost manipulators with limited mechanical repeatability. To address this issue, this paper proposes a novel Dynamic Programming–based Action Chunking Transformer (DP-ACT) framework. By integrating dynamic programming into the ACT inference chain, the method provides local trajectory correction while maintaining the implemented control-loop frequency. The proposed framework was implemented on a low-cost SO100 robotic manipulator equipped with Feetech actuators and evaluated through a three-stage precision-oriented assembly task. A total of 160 teleoperated demonstrations were collected for offline training, and 100 assembly trials were conducted for evaluation. Experimental results indicate that DP-ACT increases the assembly success rate from 62% to 82% and significantly improves trajectory efficiency while maintaining comparable task completion times.