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

A research paper introduces a new framework called Dynamic Programming–based Action Chunking Transformer (DP-ACT) designed to enhance low-cost robotic assembly tasks. The Action Chunking Transformer (ACT) is a popular method that helps robots perform assembly tasks more efficiently, but struggles with reliable assembly performance, especially for low-cost manipulators with limited mechanical repeatability.

To tackle this problem, the authors propose DP-ACT, which integrates dynamic programming into the ACT inference chain. This integration allows for local trajectory correction while preserving the control-loop frequency. The framework was tested on a low-cost SO100 robotic manipulator with Feetech actuators, and evaluated through a precision-oriented assembly task.

For offline training, the researchers collected 160 teleoperated demonstrations, and for evaluation, they conducted 100 assembly trials. The results show that DP-ACT boosts the assembly success rate from 62% to 82% and improves trajectory efficiency while maintaining similar task completion times.