a fully differentiable AV stack trainable from human demonstrations