The paper introduces GenAD, a generative framework for end-to-end autonomous driving that aims to directly produce planning results from raw sensor inputs. Unlike traditional methods that follow a serial pipeline of perception, motion prediction, and planning, GenAD models autonomous driving as a generative problem, focusing on predicting how the ego car and its surroundings evolve given past scenes. The key contributions of GenAD include: 1. **Instance-Centric Scene Representation**: GenAD transforms surrounding scenes into map-aware instance tokens, capturing high-order interactions between the ego vehicle and other agents. 2. **Trajectory Prior Modeling**: A variational autoencoder (VAE) is used to learn the future trajectory distribution in a structural latent space, considering the structural prior of realistic trajectories. 3. **Latent Future Trajectory Generation**: A gated recurrent unit (GRU) is employed to model the temporal evolution of trajectories in the latent space, generating more effective future trajectories. Experiments on the nuScenes benchmark show that GenAD achieves state-of-the-art performance in vision-centric end-to-end autonomous driving with high efficiency. The framework's effectiveness is demonstrated through its ability to handle complex traffic scenarios and produce accurate and safe trajectories.The paper introduces GenAD, a generative framework for end-to-end autonomous driving that aims to directly produce planning results from raw sensor inputs. Unlike traditional methods that follow a serial pipeline of perception, motion prediction, and planning, GenAD models autonomous driving as a generative problem, focusing on predicting how the ego car and its surroundings evolve given past scenes. The key contributions of GenAD include: 1. **Instance-Centric Scene Representation**: GenAD transforms surrounding scenes into map-aware instance tokens, capturing high-order interactions between the ego vehicle and other agents. 2. **Trajectory Prior Modeling**: A variational autoencoder (VAE) is used to learn the future trajectory distribution in a structural latent space, considering the structural prior of realistic trajectories. 3. **Latent Future Trajectory Generation**: A gated recurrent unit (GRU) is employed to model the temporal evolution of trajectories in the latent space, generating more effective future trajectories. Experiments on the nuScenes benchmark show that GenAD achieves state-of-the-art performance in vision-centric end-to-end autonomous driving with high efficiency. The framework's effectiveness is demonstrated through its ability to handle complex traffic scenarios and produce accurate and safe trajectories.