Abstract: Current Vision-Language-Action (VLA) models suffer from low inference efficiency, resulting in exclusive single-step action generation, inability to use prior action experiences, and frequent failures of robotic manipulation tasks caused by single-step deviations. We target the enhancement of inference efficiency, efficient multi-step trajectory generation and trajectory-based action optimization. Built upon the OpenVLA-7B model, a Residual VQ-VAE based discrete labeling scheme is developed. Separate training of the encoder-decoder and construction of a vector codebook effectively compress the high-dimensional continuous information of robotic actions. Discrete labels greatly lower training complexity and raise inference efficiency. Leveraging generated action sequences, an optimization method for ctions is presented to exploit preceding action experiences and strengthen model stability during execution. Evaluated on the LIBERO benchmark and in real-world environments, the improved model outperforms the baseline model in both success rate and inference efficiency, with a 78.6% increase in inference efficiency, which confirms the validity of the proposed method.