# TILDE框架实现具身基础模型蒸馏，真实任务性能提升4.1倍

> 原文链接：https://robotics.hellobike.com/papers/2026-10-18-2610-25376-96028146
> 频道：papers | 更新日期：2026-10-18 | 标签：auto, arxiv, hf-daily-papers
> 核心摘要：针对具身基础模型复现成本高的问题，提出TILDE蒸馏框架与AutoDAgger自主数据聚合管线。该方法融合在策略蒸馏与教师恢复轨迹学习，无需人工监控即可收集纠错数据，在三项真实机器人任务中性能较行为克隆提升最高4.1倍。

# TILDE框架实现具身基础模型蒸馏，真实任务性能提升4.1倍

::: v-pre
> 本文由 AI 根据公开内容整理，可能有误，请以原文为准：[TILDE: Teacher-in-the-Loop Distillation from Embodied Foundation Models](https://arxiv.org/abs/2610.25376)（HF Daily Papers · arXiv 2610.25376）

## 核心内容

为解决具身基础模型（EFMs）动作预训练对海量数据与算力的依赖，研究团队提出TILDE框架，将EFM的泛化与闭环鲁棒能力迁移至轻量级策略。该方法结合了学生诱导状态下的在策略蒸馏与教师执行恢复轨迹的学习机制，旨在解决无动作预训练策略的能力获取难题。

配套提出的AutoDAgger管线利用机器人奖励模型自动监测任务进度并触发教师干预，替代传统人工监控，使恢复轨迹收集效率提升2倍以上。实验表明，TILDE统一了示范学习、在线监督与纠错行为，支持跨架构蒸馏，在三个真实世界操作任务中成功率较纯行为克隆基线提升达4.1倍。

## 关键看点

- **AutoDAgger管线**：利用奖励模型自动触发教师干预采集恢复轨迹，数据收集效率较人工在环方法提升2倍以上
- **TILDE蒸馏机制**：协同在策略蒸馏与教师恢复轨迹学习，统一示范、在线监督及纠错行为三类信号源
- **实测性能增益**：在三项真实机器人操作任务中，任务成功率较行为克隆基线最高提升4.1倍
- **跨架构迁移**：验证了从大规模具身基础模型向异构轻量级策略进行知识蒸馏的有效性

## 工程与产业观察

该工作为具身智能提供了低成本复用大模型能力的工程范式。AutoDAgger通过自动化闭环数据采集解决了纠错数据稀缺痛点，显著降低了对人工遥操作的依赖；TILDE验证了跨架构蒸馏在真实物理环境中的有效性，为端侧部署轻量化操作策略提供了可落地的算法路径与数据构建参考。

## 英文摘要 (Abstract)

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. To transfer these capabilities from pretrained EFMs to policies without action pretraining, we introduce TILDE (Teacher-In-the-Loop Distillation from EFMs), which synergizes on-policy distillation at student-induced states with learning from teacher-executed recovery trajectories. We further propose AutoDAgger, an autonomous data aggregation pipeline in which a robotic reward model is utilized to monitor task progress and trigger teacher intervention, thereby collecting recovery experience without human monitoring or manual control. AutoDAgger accelerates recovery-trajectory collection by more than 2times compared with the human-in-the-loop method. By unifying learning from demonstrations, on-policy teacher supervision, and teacher's corrective behavior, TILDE improves task performance by up to 4.1times over behavior cloning across three real-world tasks and supports effective cross-architecture distillation.

Hugging Face Daily Papers 热度：8 upvotes
作者：Yuxiang Lu, Yantong Liu, Jinghua Hou, Xueli Sun, Ruihua Han, Zhe Liu, Hengshuang Zhao
:::

## 原始参考来源 (Sources)

- [HF Daily Papers · arXiv 2610.25376：TILDE: Teacher-in-the-Loop Distillation from Embodied Foundation Models](https://arxiv.org/abs/2610.25376)