Enhancement of Position Accuracy via DPB-Embedded ACT Model for Fine-Grained Manipulation

Since the 1960s, research on robots performing fine manipulation tasks has continuously advanced with the progress of electronic and information technologies. Despite some progress, challenges remain in areas such as the spatial perception capability of sensors, cumulative errors in imitation learning, and multimodal information fusion. The ACT (Action Chunking with Transformers) model has certain advantages in these areas, but it still faces shortcomings in complex state prediction and multi-scale, multimodal data processing. To address these issues, a Dynamic Position Bias (DPB) module is introduced to optimize action sequence outputs, enhance cross-scale feature capture and spatial perception, and improve the alignment and fusion of multimodal data. Experiments conducted on the Aloha simulation platform validate the effectiveness of the improved ACT model in two typical tasks, "Cube Transfer" and "Bimanual Insertion." Compared to the original ACT model and other baseline models, significant performance improvements were observed.