Tag
NVIDIA expanded its Cosmos platform with new training tools, simulation checkpoints and Cosmos Data Factory blueprints to help developers build, evaluate and deploy world models for physical AI. Partners including Porsche, Volkswagen Group, ANYbotics, Microsoft Azure, Nebius and CoreWeave are already using the blueprints to accelerate autonomous vehicle and robotics development.
The paper introduces SDI (Sibling Disagreement in Imagination), a method that unifies exploration and planning in world-model-based RL by using the actor's own rollout variability as a shared reference to decide when a planned action should replace the learned policy, outperforming baselines on Meta-World, Adroit, and DM-Control.
The paper proposes 'predictive closure,' a set-level characterization of how JEPA selects features to encode, showing that feature retention depends on relationships within feature sets rather than individual predictability, with experiments demonstrating that unpredictable features can be preserved when they aid prediction of others.
VPFlow argues that value prediction, not accurate state prediction, is what offline planning requires, training latent world models on closed-form Monte Carlo returns and planning in noise space via evolutionary search with a one-step flow policy. It surpasses model-free baselines by 30.3 and 10.1 percentage points on state and pixel OGBench benchmarks while outperforming latent planners like TD-MPC2.
This arXiv monograph defines physical AI and presents its technical landscape, covering multimodal perception, world models, embodied foundation models, motor control, multi-agent coordination, and hardware-software co-design, while distinguishing it from embodied AI and cyber-physical systems and outlining challenges toward Physical AGI.
ALPS is a pretrained ViT-B/16 visual state encoder for world models that unifies physical-state prediction, language alignment, and pixel reconstruction in a single patch-token representation, achieving state-probe performance matching the 3.5x larger V-JEPA 2-L and 85.5% mean policy success on LIBERO when paired with its native text tower.
The paper introduces Future Chunk Prediction (FCP), an auxiliary objective for autoregressive video diffusion models that predicts temporally separated chunks without intervening visual context, improving physical plausibility and reducing training compute. The approach requires no additional pretrained visual encoder and hints at more structured representation spaces useful for autoregressive world models.
The paper introduces AW-0, an agentic world action model framework that unifies code-based planning with joint video and action prediction in a mixture-of-transformers architecture, enabling reusable skill libraries for robot manipulation. Evaluations on RoboTwin2.0 and bimanual manipulation tasks show improved performance over VLA and WAM baselines.
ConfigWM introduces configurable perception for world models, using task descriptions to steer visual embeddings into task-specific state spaces that filter out irrelevant distractions, achieving 22.3% better planning success over the strongest baseline in noisy environments.
OverSeer introduces a multi-agent world model that augments egocentric agent views with fixed auxiliary camera views acting as a world-centric representation, compressed via delta-VAE to cut token computation, showing improved consistency in Minecraft and Doom.
Slot-JEPA 将 JEPA 的稠密表示分解为可复用的语义槽与空间掩码，提供面向对象的预测接口，并结合 Transition Memory Predictor 与非平衡最优传输提升视觉理解迁移和目标条件规划，在七项视觉基准与 REALM-JEPA 机器人规划任务上取得改进。
Matrix-Game 4.0 将游戏逻辑与视觉合成分离，通过可执行引擎与自回归视频渲染器实现低延迟的实时可玩游戏世界，在单张 RTX5090 上达到 480p、60 FPS 及 112ms 按键到显示延迟，并开源代码、模型权重与低多边形游戏资产。
Beacon shares its research vision for Company World Models (CWM): AI systems that observe business activity, hypothesize and implement strategic improvements by dispatching digital employees, and measure outcomes to drive company-level decision making. The post outlines the capability and alignment challenges involved and argues Beacon's structure enables the fastest research path toward this goal.
Tsinghua and Didi researchers propose the World Action Critic (WAC) Model, which jointly learns future prediction and consequence evaluation for counterfactual action candidates in world-model-based policy learning, achieving 93.4 PDMS on the NAVSIM-v1 autonomous driving benchmark.
GlobeBench 提出一个包含 1,632 个任务的标准化基准，用于评估文本世界模型模拟 RL 环境状态转移的能力；评测 20 个 LLM 后发现即使是表现最好的 GPT-6 Astra 任务成功率也仅 49.57%，且 631 个任务无任何模型解决，凸显模型作为通用世界模型仍存在显著差距。
WildAWM 提出一个可扩展的流水线，用于合成可控的计算机使用环境与任务来训练通用智能体，支持 GUI、MCP 和 CLI 交互，并在 8 个分布外基准上取得显著提升。
Ropedia W0 是首个跨人类与多机器人具身的 loco-manipulation 世界基础模型，通过统一运动学表征和 10k 小时人类 + 3k 小时机器人数据实现五个基准上的最先进性能，蒸馏后支持实时分钟级 rollout 与 VR 交互。
研究人员提出 AnyGoal，一个基于目标偏移任务变体的世界模型基准与开源库，用于评估世界模型能否将熟悉的状态与动作重新组合成有效动力学。实验发现现有世界模型家族在这些重组场景下普遍失效，揭示了动力学泛化与规划性能之间的关键差距。
EgoXperience 提出一个统一框架，通过自动标注的大规模数据和单次前向传播模型，从第一人称视频中联合重建相机运动、场景几何、手部与物体，并发布了 EgoX-Data-6M 数据集与 EgoX-Bench 基准，取得显著优于基线的性能。
The paper introduces Stateful Harness, a framework that separates rule-based world state updates from frozen video rendering to improve consistency in interactive world models with hidden state, and evaluates it with the WorldState-Bench benchmark, where the full harness raises rule-compliance from 32% to 92%.