#world models

437 stories · 37 in 7 days · 124 in 30 days
Nov 13
Papers10:55
Chinese content论文揭示世界模型想象 rollout 会逐步丢失任务信号

做 model-based RL 或 active inference 规划的人该看看。它用 CARLA 里的探针实验证明想象 rollout 会悄悄丢掉障碍物信息,而且 ensemble 不确定性完全察觉不到,0.52 的 AUC 就是随机水平。这种可测量的诊断思路比单纯刷分有用。

arXiv cs.AI@Jaerock Kwon et al. (4)Source
ProductsBuildersOfficial10:36
Chinese contentACE Robotics 用 Mundo 世界模型让机器人完成酒店洗衣全流程

ACE Robotics 把 Mundo 世界模型和机上技能拼成了一套酒店洗衣机器人,失败案例能反哺技能改进,单轮 18.6 秒。想看世界模型怎么在真实场景落地的,这篇比纯实验室 demo 有参考价值,做具身智能的可以细看。

Primary sourcePandaily@[email protected] (Pandaily)Source
ModelsBuildersOfficial10:34
Chinese content华为云 11 月 16 日发布 CloudRobo-WM1.2:首个原生基于昇腾的世界模型

华为云把世界模型和自家昇腾芯片绑在一起做了,不依赖英伟达生态。做机器人训练、具身智能的团队可以关注,毕竟全昇腾链路的具身模型目前国内还真不多见。

Event
Primary sourcepandaily@[email protected] (Pandaily)Source
ProductsBuildersMulti-source72°02:57
Chinese contentWorld Labs 向所有用户开放 3D 世界模型 Atlas

李飞飞创办的 World Labs 把 Atlas 世界模型全量开放了,浏览器里就能用,API 也一并上线。它输出的不只是视频,还有能直接进 3D 渲染管线的高斯泼溅资产,和 Sora、Veo 这类纯视频产品思路不同。做游戏场景、虚拟空间或仿真数据的开发者值得上手试试。

Event
Official accountFei-Fei Li@drfeifeiSource
Nov 12
Chinese content华为云 CloudRobo-WM1.2 昇腾原生世界模型 11 月 16 日发布

华为云要出手了,11 月 16 日发布首款昇腾原生的世界模型 CloudRobo-WM1.2。做机器人开发的值得看直播,华为云明确不造本体、只做平台,等于给国内一堆机器人公司提供了底层模型选项,跟英伟达 Cosmos 这类世界模型形成了对标。

Event
IT之家Source
Chinese contentPAN 2 开源:动作条件世界模型,可生成最长 20 秒结果预测视频

做 Agent 或机器人训练的朋友可以关注:PAN 2 开源了,输入一个动作描述,语言模型先推演剧情,29B 视频模型再渲染出最长 20 秒的物理结果,等于把试错搬进了仿真,不用在真实环境里烧成本。

rohanpaul_ai@rohanpaul_aiSource
Chinese contentIFM 开源世界模型 PAN 2,用模拟式推理让智能体先演练再行动

IFM 把世界模型做成了智能体的"预演器":29B 的 LDM 负责画后果,LLM 负责想后果,选最优再动手。和纯文本推理的 agent 框架相比,它多了视频级的可视化验证。做机器人、游戏或决策型智能体的开发者可以直接下载权重和代码上手试。

Ate-a-Pi@svpinoSource
Nov 11
ModelsBuildersOfficial15:44
Chinese contentMBZUAI 旗下 IFM 发布开源世界模型 PAN 2:29B 参数可模拟 20 秒动作结果

阿联酋 MBZUAI 的 IFM 实验室开源了 PAN 2,用 29B 扩散模型加 LLM 规划器,能预演最长 20 秒的物理动作结果。做机器人或具身智能的团队可以自己部署跑起来,不用从零搭仿真环境,Apache 2.0 许可商用也没障碍。

Primary sourcemarktechpost@Michal SutterSource
ModelsBuildersMulti-source14:42
Chinese contentFable 5.2 演示扩展到整片城市:伊斯坦布尔、恐龙世界与浏览器开放世界

Fable 5.2 的 5 个新演示挺直观的:有整座伊斯坦布尔、能打碎的体素场景,还有一个直接开浏览器就能玩的开放世界。做游戏或世界模型方向的人可以看看这批案例,感受下这类模型和静态视频生成的差别。

Event
AI Will@FinanceYF5Source
Chinese content成都人形机器人创新中心发布第五代贡嘎人形机器人

成都人形机器人创新中心的第五代贡嘎把记忆、因果推理和世界模型整合成了一套完整链路,机器人能记住物品位置并跨越多天完成任务,遇到意外还能自主改方案。做具身智能或家庭服务机器人的团队可以看看它的技术架构思路。

IT之家Source
ModelsBuildersSelected02:26
MBZUAI's IFM Releases PAN 2 Open World Model

The ablation numbers reveal that planning, not visual fidelity, is the bottleneck in multi-step world-model tasks, a finding that will shape how teams architect embodied agents. Robotics researchers and agent builders can use the open weights to benchmark simulative reasoning pipelines against proprietary alternatives. As a UAE-based lab's release, it adds a non-US, non-China option to the open world-model landscape, relevant to ASEAN academic teams seeking accessible baselines.

rohanpaul_ai@rohanpaul_aiSource
Chinese contentPAN 2 开源世界模型发布,vLLM-Omni 提供 Day-0 推理支持

PAN 2 团队开源了一个按时间、物理、空间来预测场景变化的世界模型,能做文生视频和图生视频。vLLM-Omni 第一天就支持它,想在自己服务器上跑视频生成或做真实世界模拟的开发者可以直接用开源推理栈部署了。

Official accountvLLM@vllm_projectSource
IndustryPolicySelected00:55
Sakana AI's Schmidhuber says Physical AI needs new data

The interview offers a counterweight to hype cycles: a foundational researcher arguing embodiment and self-generated data, not more web-scale text, define the next phase. Robotics teams and enterprise planners evaluating Physical AI roadmaps get a concrete, sceptical timeline from a credible source. The framing of Japan and Germany as potential contenders in embodied AI gives readers outside the US-China race a lens on where the field may diversify.

Official accountSakana AI@SakanaAILabsSource
Nov 10
Chinese contentLibrary of Hands:免训练修正世界模型的动作映射,机器人成功率从 28.6% 提到 62.3%

搞机器人世界模型的应该看看这篇。相机一挪、控制器一换,模型就得重训,这活儿很烦。LoH 完全不训练,靠前后两帧图像反推出实际动作映射,还能把映射存起来复用。成功率直接翻倍,开销几乎为零,工程上很容易试。

arXiv cs.AI@Hasaan Ahmad et al. (5)Source
Nov 9
PapersBuildersSelected71°05:07
Yann LeCun Clarifies the Difference Between JEPA and JEA

The real takeaway is a taxonomy: JEA methods like DINO and Barlow Twins suit symmetric view-matching, while JEPA's predictor targets corrupted-input or temporal prediction, with an action variable making it a world model. Researchers comparing self-supervised approaches or building world models get a concise map of which loss function fits which task.

Official accountYann LeCun@ylecunSource
Nov 6
ModelsBuildersSelected11:29
PlayGen2 Real-Time World Model Uses Explicit 3D State

The architectural split between 3D state prediction and rendering avoids the drift that plagues video-frame-recurrent world models, and the laptop-scale latency shows the approach runs on consumer hardware rather than data-centre GPUs. Developers building playable-world or interactive-simulation demos can download the code and weights and extend the framework to other games or rule-based gameplay. It compares with video-only world models like Genie-style systems by trading appearance realism for spatial persistence and control.

arXiv cs.LG@Jihai Zhang et al. (6)Source
ModelsBuildersSelected02:50
Verne Robotics V0.5 open-weight world action model

Robustness to scene changes, not raw benchmark scores, is what separates deployable robot models from lab demos, and V0.5 posts a much smaller drop than π0.5 under distribution shift. Robotics teams evaluating open-weight manipulation models can benchmark V0.5 against π0.5 in their own cluttered environments. Its base on the Wan2.2 video generator also shows video world models doubling as action predictors.

rohanpaul_ai@rohanpaul_aiSource
Nov 5
Nov 4
Nov 3
ProductsBuildersSelected21:01
Baidu Apollo World World Model for Autonomous Driving Debuts

World models are becoming the training ground where driving policies are stress-tested before touching real roads, and Baidu's 2+3 criteria give the field a concrete checklist for judging them. Engineers working on autonomous driving, robotics or embodied AI can use the framework to benchmark competing world models. Baidu (Apollo) also signals how Chinese AV players are closing the sim-to-real loop domestically, a useful comparison point against Waymo and Tesla approaches.

Event
IT之家Source