PAVXploreRL: Action-Exploration RL Framework Targets Physical-Action-Visual World ModelsPAVXploreRL：动作探索强化学习框架瞄准物理-动作-视觉世界模型
S 1.7T12 sources2 个来源R7-research
Han Wang and six co-authors (including Xiaodan Liang and Roy Ka-Wei Lee) propose PAVXploreRL (arXiv:2607.16602v2), a reinforcement-learning framework for action-conditioned world models—used as scalable policy evaluators to cut reliance on costly real-world rollouts—built around three target properties dubbed "PAV": Physical Plausibility, Action Adherence, and Visual Fidelity.
The paper, first submitted 18 Jul 2026 and revised 21 Jul 2026, states that existing PAV world models struggle to stay robust across both in-distribution (ID) expert demonstrations and out-of-distribution (OOD) actions, which motivates PAVXploreRL's added action-exploration mechanism named in its title.
Because such world models stand in for expensive real-world rollouts when evaluating policies, the PAV framework bears directly on autonomous driving, where cheaply and safely simulating rare or OOD driving actions remains a key bottleneck for AV validation.
Han Wang联合六位共同作者（包括梁小丹和Roy Ka-Wei Lee）提出PAVXploreRL（arXiv:2607.16602v2），这是一个面向动作条件世界模型的强化学习框架——该类模型可作为可扩展的策略评估器，减少对昂贵真实世界数据采集的依赖——并围绕物理合理性（Physical Plausibility）、动作粘性（Action Adherence）、视觉保真度（Visual Fidelity）三项目标（合称PAV）展开。