TILDE Distills Embodied Foundation Models via Teacher-in-the-Loop
October 17, 2026
TILDE enables the transfer of manipulation capabilities from large Embodied Foundation Models to smaller policies without requiring action pretraining. It uses an AutoDAgger pipeline and a robotic reward model to automate the collection of recovery trajectories.
HOW THIS AFFECTS YOU
●
builderYou can train robust robotic policies without the massive compute required for full action pretraining.
●
researcherThe AutoDAgger method provides a way to automate recovery data collection without human monitoring.