TILDE distills embodied foundation models without action pretraining
Original title: TILDE: Teacher-in-the-Loop Distillation from Embodied Foundation Models
AutoDAgger collected recovery trajectories more than twice as fast as a human-in-the-loop method.
TILDE transfers manipulation skills from pretrained embodied foundation models to student policies that skip large-scale action pretraining. It combines on-policy distillation at student-induced states with learning from teacher recovery trajectories. AutoDAgger uses a robotic reward model to trigger teacher intervention without human monitoring.
Unconfirmed: The reported speed and performance multiples are author claims without independent confirmation in the text.
- Authors report up to a 4.1-times gain over behavior cloning.
- The comparison covers three real-world tasks.
- The method is said to support cross-architecture distillation.
- AutoDAgger monitors task progress and collects recovery experience autonomously.
From the original article
1 min read at huggingface.co
Large-scale action pretraining equips Embodied Foundation Models (EFMs) with strong manipulation capabilities, with particular strengths in generalization and closed-loop robustness, but reproducing such pretraining requires substantial robot data and computation.
By Yuxiang Lu, Yantong Liu, Jinghua Hou · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.