# 具身智能CPT研究：数据多样性优于模型规模提升跨世界泛化

> 原文链接：https://robotics.hellobike.com/papers/2026-10-20-2610-28698-4327a145
> 频道：papers | 更新日期：2026-10-20 | 标签：auto, arxiv, hf-daily-papers
> 核心摘要：研究通过120组配置验证具身智能持续预训练缩放规律，发现数据多样性显著提升相关世界迁移能力，而单纯扩大模型规模在远距离世界中泛化反而下降，揭示了构建通用具身智能体需超越单一CPT阶段。

# 具身智能CPT研究：数据多样性优于模型规模提升跨世界泛化

::: v-pre
> 本文由 AI 根据公开内容整理，可能有误，请以原文为准：[Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds](https://arxiv.org/abs/2610.28698)（HF Daily Papers · arXiv 2610.28698）

## 核心内容

针对虚拟环境中具身学习是否遵循物理世界缩放规律的问题，研究团队开展了涵盖数据量、多样性与模型容量三轴线的持续预训练（CPT）对照实验。测试覆盖0.8B至27B四种模型规模、50至1600小时六档数据预算及五种数据混合策略，累计完成120组训练配置，并通过三级验证损失与跨世界游戏基准评估性能。

实验证实增加数据量可提升语料内性能，数据多样性有效改善向相关世界的迁移，但大模型在单世界训练下对远距离世界泛化较差。研究识别出早期“接口阶段”具备广泛跨世界收益，后期“世界阶段”增益收窄，并拟合了验证损失与三要素的经验缩放律。模型在3D实时场景中表现优异，但模仿学习局限表明通用具身智能仍需更丰富训练范式。

## 关键看点

- **缩放规律验证**：120组CPT配置证实数据多样性比模型扩容更能促进跨世界迁移，大模型在单世界训练下远距离泛化反降
- **训练阶段分化**：发现早期接口阶段具广泛跨世界收益，后期世界阶段增益局限于语料内及相关世界，并拟合经验缩放律
- **实时性能达标**：模型在Minecraft等3D世界实时设置中表现最佳，竞争力对标专用模型与前沿大模型
- **通用性瓶颈**：单纯模仿学习与CPT在远距离世界存在持久差距，构建通用具身智能体需引入更丰富的训练阶段

## 工程与产业观察

该研究为具身基础模型训练提供了关键工程坐标：在算力受限时，优先构建多世界异构数据集比盲目堆叠参数更具性价比。发现的“接口-世界”两阶段特征提示开发者，CPT仅能解决基础感知运动对齐，实现真正跨域泛化必须叠加任务规划或强化学习等后训练阶段，避免陷入模仿学习的局部最优陷阱。

## 英文摘要 (Abstract)

Scaling data, diversity and model capacity has improved embodied learning in physical environments, where diverse experiences remain grounded in shared physical laws. Whether these gains extend to virtual worlds with different visual styles, dynamics, controls, and objectives remains unclear. We investigate this question through a controlled study of continued pre-training (CPT) across three axes: data volume, diversity, and model size. Our 120 training configurations span four model sizes from 0.8B to 27B, six data budgets from 50 to 1,600 hours, and five mixtures ranging from Minecraft-only data to diverse multi-world corpora. We assess performance and generalization using three-level validation losses and gameplay benchmarks from in-corpus worlds to unseen worlds. Our results show: (1) more data improves in-corpus performance, diversity improves transfer to related worlds with limited gains in distant worlds, and model size improves data efficiency but generalizes worse to distant worlds under single-world training; (2) we observe an early interface phase with broad cross-world benefits and a later world phase with gains narrowed to in-corpus and related worlds, and fit an empirical scaling law relating validation loss to data volume, diversity and model size; and (3) our models are competitive with Minecraft specialists and frontier models in 3D worlds, and perform best in real-time settings. However, the limits of imitation and persistent gaps in distant worlds suggest that building generalist embodied agents requires richer training stages beyond continued pre-training.

Hugging Face Daily Papers 热度：20 upvotes
作者：Kuan Zhang, Yukun Chen, Zhihao Yang, Yue Su, Xiangnan Wu, Zirong Chen, Run Luo, Jinkun Hou, Tao Tan, Yinhe Zheng, Yiming Li
:::

## 原始参考来源 (Sources)

- [HF Daily Papers · arXiv 2610.28698：Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds](https://arxiv.org/abs/2610.28698)