# 通过模型到策略表述统一具身基础模型

> 原文链接：https://robotics.hellobike.com/papers/2026-10-16-2610-23053
> 频道：papers | 更新日期：2026-10-16 | 标签：arxiv, cs.RO
> 核心摘要：针对 VLA 与 WAM 的架构和目标差异，提出基于随机核的架构无关概率表述，并在 OpenVLA 和 DreamZero 上实例化。

# 通过模型到策略表述统一具身基础模型

::: v-pre
> 原文链接：[Unifying Embodied Foundation Models via a Model-to-Policy Formulation](https://arxiv.org/abs/2610.23053)（arXiv 2610.23053 · 发布日期：2026-10-16）

**Unifying Embodied Foundation Models via a Model-to-Policy Formulation**

针对 VLA 与 WAM 的架构和目标差异，提出基于随机核的架构无关概率表述，并在 OpenVLA 和 DreamZero 上实例化。

## 核心要点

- 动机是统一 VLA 模型与 WAM 在机器人学习和控制中的不同架构、输出表示与学习目标。
- 方法用条件输出分布和动作实现核刻画具身基础模型，从而诱导任务条件机器人策略。
- 框架将预训练和后训练连接到策略部署，兼容特定模型预训练目标及离线或在线后训练。
- 论文在 OpenVLA 和 DreamZero 上实例化，覆盖自回归动作词元预测与联合视频-动作生成。

## 英文摘要 (Abstract)

Vision-language-action (VLA) models and world-action models (WAMs) have emerged as promising approaches to robot learning and control. These models integrate multimodal perception and action generation, but differ in their architectures, output representations, and learning objectives. In this paper, we present a unified, architecture-agnostic probabilistic formulation for embodied foundation models based on stochastic kernels. Our formulation has two key components: (i) we characterize embodied foundation models through conditional output distributions and action-realization kernels, which together induce task-conditioned robot policies; and (ii) we connect pretraining and post-training to policy deployment, accommodating model-specific pretraining objectives and offline or online post-training strategies. We instantiate the formulation for OpenVLA and DreamZero, covering autoregressive action-token prediction and joint video-action generation.

作者：Xinyi Hong, Ren Mingyang
:::

## 原始参考来源 (Sources)

- [arXiv 2610.23053：Unifying Embodied Foundation Models via a Model-to-Policy Formulation](https://arxiv.org/abs/2610.23053)