Tag
The paper proposes Latent Interface Training (LIT), a two-stage strategy to enhance generalization in robotics foundation models by mitigating vision-action shortcuts through pose-supervised latent interfaces.