Researchers led by Haisen He have built Hindsight Self-Distillation, a training method that uses the later frames of recorded demonstrations to teach a robot policy what to pick out of its current camera images, according to a paper on arXiv. In training, a teacher sees both current and later frames, while a student that shares its encoder sees only the current ones.
The student learns from the demonstrations or the teacher's predictions, and learns to match the teacher's visual summaries. Once deployed, it encodes the current images once and reuses that summary at every step of its action generation. Its average success rate reached 98.50% on the LIBERO benchmark and 92.31% on RoboTwin 2.0.



