跨实体迁移中的配对数据与多样性之争

⌘K

Change language Switch ThemeSign In

Narrow Mode

Paired Data vs. Diversity in Cross-Embodiment Transfer ======================================================

Paired Data vs. Diversity in Cross-Embodiment Transfer ======================================================  ### Chelsea Finn

@chelseabfinn

通常，我们认为数据多样性越高越好。在跨实体迁移中，来自不同实体的配对数据似乎比单纯增加多样性更能带来益处。

网页和代码：data-analogies.github.io

论文：arxiv.org/abs/2603.06450

!Image 3: Tweet image

Mar 16, 2026, 12:26 AM View on X

6 Replies

30 Retweets

225 Likes

11.1K Views  Chelsea Finn @chelseabfinn

One Sentence Summary

Chelsea Finn shares research suggesting that paired data across different robot embodiments is more effective for transfer learning than simply increasing data diversity.

Summary

This tweet highlights a key finding from a new research paper regarding robotics and cross-embodiment transfer. While the common intuition in machine learning is that more diverse data leads to better generalization, this study indicates that for transferring skills between different types of robots (embodiments), having 'paired data' (analogous actions or states across different robots) provides a more significant benefit than diversity alone. The tweet includes links to the project website, code, and the full paper on arXiv for further technical exploration.

AI Score

81

Influence Score 66

Published At Today

Language

English

Tags

Cross-Embodiment Transfer

Robotics

Transfer Learning

Paired Data

Machine Learning Research HomeArticlesPodcastsVideosTweets

Paired Data vs. Diversity in Cross-Embodiment Transfer | ... ===============