Abstract: To improve the efficiency and accuracy of robot motion skill learning, in this paper, a novel action decision Transformer model based on sequential feature processing, named GMT (gated mechanism Transformer), was proposed. Using GPT-2 as its core architecture and incorporating a gating mechanism to extract hidden state features, the model addressed the challenge of extracting deep features from robot motion data by capturing temporal dependencies through autoregressive modeling. Meanwhile, the model employed parameter sharing strategies to refine predictive features for action inference. GMT was evaluated on three robot motion skill tasks on the MuJoCo platform. Experimental results demonstrate that, compared with Decision Transformer, GMT achieved up to 28.5% improvement in learning efficiency and accuracy. The research indicates that GMT can effectively model robot motion sequence features, providing a new technical approach for robot action decision-making.