GenAD: How to implement trajectory prior modeling to improve autonomous driving performance?

The GenAD project uses a variational autoencoder to model the distribution of future intentions. Compared with the traditional method of modeling multiple trajectories separately, this method uses distribution to represent future trajectories, which is more reasonable and efficient. The distribution in the latent space is represented by Z, and by sampling in this space, a specific trajectory can be obtained. In order to learn this Z distribution, the project takes the ground truth trajectory as input and maps it to the latent space through an encoder.