Four stages of autonomous driving architecture evolution

74
At present, the development of end-to-end autonomous driving architecture can be divided into four main stages. The first stage is "end-to-end" perception, which improves the accuracy and stability of perception output through multi-sensor fusion technology and transformer networks. The second stage is decision-making and planning modeling, integrating prediction, decision-making and planning into a neural network. The third stage is modular end-to-end, where the perception module outputs feature vectors, and the decision-making and planning module outputs results based on the feature vectors and is trained through gradient conduction. The fourth stage is One Model end-to-end, using a deep learning model to process the entire process from raw signals to planned trajectories.