Zhenjie Zhu

Robot policies can achieve faster task completion by learning to decouple *where* to go from *when* to execute, outperforming fixed-interval action prediction.

APT achieves substantial improvements in instruction generalization for VLA models by effectively decoupling language and action learning, addressing a critical data imbalance issue.

Forget painstakingly tuning force magnitudes for sim-to-real transfer of contact-rich tasks; this work shows that learning just the *direction* of contact forces unlocks surprisingly robust performance.