Abstract
In the field of robotics control, prevailing research is progressively leveraging more sophisticated deep learning networks to enhance learning outcomes in specific domains of robotics motion control. Notably, the Decision Transformer (DT), a prominent model in motion control learning, predominantly focuses on long-term expected returns and often overlooks the critical significance of immediate action values, which are essential for the adaptive behaviors observed in higher animals. Addressing this gap, this paper introduces the Focused Reward Transformer (FRT) . This innovative training network combines guided attention mechanisms with a significant emphasis on reward processing. The FRT aims to optimize robot motion control by dynamically balancing long-term predictions with immediate environmental cues. The effectiveness of the FRT was rigorously evaluated by detailed simulations on the MuJoCo platform, using three different tasks: Hopper, Walker2d, and Halfcheetah. These tasks broadly represent the challenges inherent to robotics dynamics. Our experimental results demonstrate a substantial improvement in the performance of control tasks, the ability to validate FRT, refine robotic control, and skillfully combine in-time context responses with the long-term planning of strategies. This confirms the potential of FRT to advance the field of robotic control by introducing more nuanced approaches to learning and decision-making processes.