Few-shot decision-making method for high-speed driving based on large model architecture

Compared with low-speed automatic driving, high-speed automatic driving often leads to safety accidents. This paper focuses on the decision-making of high-speed safe driving control, and combines the characteristics of expressway automatic driving, and puts forward a two-level bidirectional iterative algorithm. The real-time continuous control decision is modelled as an action sequence, and the action sequence is fitted through the transformer architecture. On this basis, this paper designs a novel two-layer learning algorithm to mine the context information between periodic sequence nodes and the association information between action sequence nodes. The system model and novel algorithm show fast convergence speed, strong Few-shot training ability and adaptability to different environments in the automatic driving environment of expressway. The results show that a large number of calculation problems caused by continuous control can be alleviated by establishing action sequence and periodic sequence. At the same time, through the combination of meta-learning and Transformer architecture, as well as the two-way iteration of the inner and outer layers, a better large model paradigm can be formed, and the problem of environmental adaptability can be solved by using Few-shots.