外观
论文arXiv 2609.166971 个月前
面向具身智能的世界模型：从可信到可控再到可行动
原文链接：World Models for Embodied Intelligence: From Plausible to Controllable to Actionable（arXiv 2609.16697 · 发布日期：2026-09-11）
World Models for Embodied Intelligence: From Plausible to Controllable to Actionable
本文针对具身智能世界模型如何改善行为的问题，提出三级能力框架和3x4分析矩阵，并在多类具身任务文献中梳理。
核心要点
- 动机是评估不应只看视觉逼真度，而应关注预测如何支持干预、决策与闭环行为改进。
- 方法上将世界模型能力划分为可信、可控、可行动三级，并结合几何、物理和动作落地等维度分析。
- 综述覆盖操作、导航、运动、自主驾驶和通用具身学习，梳理方法进展、数据集、基准与评估协议。
- 指出开放挑战包括长时域一致性、不确定性校准、因果干预测试、延迟、验证恢复和跨具身迁移。
验证范围与局限：验证范围为文献综述，覆盖操作、导航、运动、自主驾驶和通用具身学习，摘要未报告新的实机或仿真实验。
英文摘要 (Abstract)
World models are emerging as a key interface between perception and decision making in embodied intelligence, enabling agents to preserve hidden state, anticipate consequences, compare interventions, and adapt when execution departs from expectation. Much of this progress has been measured by visual fidelity. For an embodied agent, however, the purpose of a world model is to improve behavior by predicting futures that inform intervention and decision making. Before reaching for a cup, a person anticipates its weight and how it will resist the grasp, and the hand shapes itself before contact. That anticipation is coarse and rarely pictorial, yet it guides the movement. The key question is therefore which properties of a prediction let it improve behavior. Existing surveys organize this literature by architecture, output modality, or application domain, which leaves this question implicit. We accordingly characterize world model capability through three progressively stronger levels: A Plausible world model preserves task-relevant temporal, geometric, or physical structure; a Controllable model additionally predicts how interventions change that structure; and an Actionable model uses prediction to produce measurable improvements in planning, action, learning, evaluation, verification, recovery, or data selection. We complement this capability ladder with a 3x4 analysis matrix that crosses geometry, physics, and action grounding with data, reward, policy, and model-self improvement loops. Using this framework, we survey research across manipulation, navigation, locomotion, autonomous driving, and general embodied learning. We organize methods as technical progressions, clarify the capability required at each level, and examine datasets, benchmarks, and evaluation protocols. Finally, we identify open challenges in long-horizon consistency, uncertainty calibration, causal intervention testing, latency, verification and recovery, and cross-embodiment transfer. This perspective motivates a shift in evaluation from whether a model can generate a convincing future to whether it represents the right state, responds correctly to action, and improves the closed-loop behavior of an embodied agent.
作者：Nanjie Yao, Hao Wang, Chong Cheng, Zhikang Chen, Wenzhe Li, Jiafei Lyu, Li Shen, Peilin Zhao, Zongqing Lu, Gao Huang
来源
本页事实对应的公开页面，以原文为准。