A new preprint on arXiv introduces TILDE (Teacher-in-the-Loop Distillation from Embodied Foundation Models). The abstract highlights that large-scale action pretraining gives Embodied Foundation Models (EFMs) strong manipulation capabilities, particularly in generalization and closed-loop robustness. These are valuable traits for real-world robotic tasks.

The paper then notes that reproducing such pretraining is difficult, though the abstract's sentence is cut off at that point. TILDE is presented as a method to distill from these large models, presumably into a smaller or more efficient student, using a teacher-in-the-loop framework. The exact mechanics of the distillation are not detailed in the available abstract.

As the source is only the abstract, specifics about the architecture, training data, or experimental results are not provided. The key contribution, based on the text, is the proposal of a distillation approach that leverages a teacher model during training to overcome the barriers to reproducing large-scale action pretraining.