We find that neural collapse and concept alignment enable learning interpretable control vectors, which we refine with sparse autoencoders.

An end-to-end autonomous driving approach that separates semantic and motion learning into parallel decoders to mitigate negative transfer.

JointMotion connects scene-level motion and environment embeddings via a non-contrastive alignment objective, then applies masked polyline modeling to unify global context and instance-level representation.

RedMotion fuses local road features into a global embedding via an internal decoder, then applies self-supervised redundancy reduction across augmented views to unify local and global road representations.