GenAD: How to achieve efficient autonomous driving motion prediction and planning?

The GenAD project innovatively uses variational autoencoders to learn the distribution of future trajectories in the structured latent space for trajectory prior modeling. In addition, the project also uses temporal models such as gated recurrent units to capture the subject and self-motion in the latent space to generate more effective future trajectories. This forward-looking approach enables GenAD to sample distributions from the conditioned structured latent space during reasoning and use the learned temporal model to generate the future, achieving simultaneous execution of motion prediction and planning.