Decoupled Reading
Letting demonstration-conditioned robot policies follow longer compositions
Demonstration-conditioned policies learn to imitate a task from a prompt demonstration, but struggle when that demonstration chains together more sub-tasks than the compositions seen in training. We decouple reading the demonstration from acting: a small reader (~0.4M parameters), trained only on short compositions with DTW-aligned pointer labels, tracks where the robot is in a long demonstration and feeds a frozen policy a moving read window — no policy updates required. Joint work with Cheng Yang and Prof. Jiatao Gu. Accepted to the CoRL 2026 Workshop on Compositional Generalization for Real-World Robot Learning.
