[
  {
    "id": "world-models-2018-release",
    "project_id": "world-models-2018",
    "date": "2018",
    "event": "The World Models project published its latent-model-and-controller approach.",
    "sources": [
      "https://worldmodels.github.io/"
    ],
    "imported_on": "2026-07-09",
    "title": "World Models 提出梦境中的智能体训练",
    "summary": "在世界模型生成的梦境环境中训练智能体，再将策略迁回实际环境。"
  },
  {
    "id": "daydreamer-corl-2022",
    "project_id": "daydreamer",
    "date": "2022",
    "event": "DayDreamer presented world-model-based online learning on physical robots at CoRL 2022.",
    "sources": [
      "https://danijar.com/project/daydreamer/"
    ],
    "imported_on": "2026-10-01",
    "title": "DayDreamer 将世界模型用于真实机器人学习",
    "summary": "DayDreamer 研究在四种真实机器人上直接在线学习，通过学得的世界模型想象未来并改进策略，不依赖仿真器。"
  },
  {
    "id": "copilot4d-2024-iclr",
    "project_id": "copilot4d",
    "date": "2024",
    "event": "Copilot4D was presented as an ICLR 2024 world model for lidar forecasting.",
    "sources": [
      "https://waabi.ai/research/copilot-4d"
    ],
    "imported_on": "2026-07-27",
    "title": "Copilot4D 登上 ICLR 2024",
    "summary": "先将传感器观测编码为 token，再通过离散扩散预测未来的世界模型。"
  },
  {
    "id": "diamond-atari-neurips-2024",
    "project_id": "diamond-atari",
    "date": "2024",
    "event": "DIAMOND studied reinforcement learning in a diffusion world model for Atari at NeurIPS 2024.",
    "sources": [
      "https://arxiv.org/abs/2405.12399"
    ],
    "imported_on": "2026-10-01",
    "title": "DIAMOND 研究用扩散世界模型训练 Atari 智能体",
    "summary": "DIAMOND 根据历史画面和动作生成后续游戏画面，让强化学习智能体在扩散世界模型中训练。"
  },
  {
    "id": "sora-world-simulators-2024-note",
    "project_id": "sora-world-simulators",
    "date": "2024",
    "event": "OpenAI published Video Generation Models as World Simulators.",
    "sources": [
      "https://openai.com/index/video-generation-models-as-world-simulators/"
    ],
    "imported_on": "2026-07-09",
    "title": "OpenAI 发布 Sora 世界模拟器技术报告",
    "summary": "OpenAI 在这篇技术报告中介绍可生成时长可变、最长一分钟视频的 Sora。"
  },
  {
    "id": "vista-2024-weights",
    "project_id": "vista-2024",
    "date": "2024-06-06",
    "event": "Vista released its v1.0 driving-world-model weights on 6 June 2024.",
    "sources": [
      "https://github.com/OpenDriveLab/Vista"
    ],
    "imported_on": "2026-10-01",
    "title": "Vista 公开驾驶世界模型权重",
    "summary": "Vista 官方仓库宣布公开 v1.0 模型权重，为驾驶场景生成与可控未来模拟研究提供下载入口。"
  },
  {
    "id": "genie-3-2025-release",
    "project_id": "genie-3",
    "date": "2025",
    "event": "Google DeepMind introduced Genie 3 as a general-purpose world model for real-time navigable generated worlds.",
    "sources": [
      "https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/"
    ],
    "imported_on": "2026-07-09",
    "title": "Google DeepMind 介绍 Genie 3",
    "summary": "根据文本提示生成可实时导航的动态世界。"
  },
  {
    "id": "vjepa-2-ac-2025-paper",
    "project_id": "vjepa-2-ac",
    "date": "2025",
    "event": "The V-JEPA 2 v1 paper described V-JEPA 2-AC as a separately action-conditioned robot-planning route.",
    "sources": [
      "https://arxiv.org/abs/2506.09985v1",
      "https://ai.meta.com/blog/v-jepa-2-world-model-benchmarks/"
    ],
    "imported_on": "2026-07-12",
    "title": "V-JEPA 2 论文介绍动作条件变体",
    "summary": "冻结预训练视频编码器后，在其表征之上另行训练的动作条件预测器。"
  },
  {
    "id": "wham-nature-2025",
    "project_id": "wham-2025",
    "date": "2025-02-19",
    "event": "The Nature paper on World and Human Action Models was published on 19 February 2025.",
    "sources": [
      "https://www.nature.com/articles/s41586-025-08600-3"
    ],
    "imported_on": "2026-10-01",
    "title": "WHAM 论文发表：预测游戏画面与玩家动作",
    "summary": "Nature 论文介绍 WHAM，从人类游玩数据学习游戏画面与控制器动作，可通过画面或动作条件探索游戏场景。"
  },
  {
    "id": "dreamerv3-nature-2025",
    "project_id": "dreamerv3",
    "date": "2025-04-02",
    "event": "The Nature paper on DreamerV3 was published on 2 April 2025.",
    "sources": [
      "https://www.nature.com/articles/s41586-025-08744-2"
    ],
    "imported_on": "2026-10-01",
    "title": "DreamerV3 论文发表于 Nature",
    "summary": "DreamerV3 学习预测动作后果的世界模型，并在想象轨迹中改进行为；论文研究多种控制任务中的学习方法。"
  },
  {
    "id": "dino-wm-icml-2025",
    "project_id": "dino-wm",
    "date": "2025-07",
    "event": "DINO-WM appeared in the ICML 2025 proceedings in July 2025.",
    "sources": [
      "https://proceedings.mlr.press/v267/zhou25t.html"
    ],
    "imported_on": "2026-10-01",
    "title": "DINO-WM 研究在预训练视觉特征上规划",
    "summary": "DINO-WM 利用 DINOv2 的空间特征和离线行为轨迹学习动作条件预测，在测试时优化动作序列以接近目标图像的特征。"
  },
  {
    "id": "marble-2025-release",
    "project_id": "marble",
    "date": "2025-11-12",
    "event": "World Labs announced Marble for creating, editing, expanding, combining, and exporting 3D worlds.",
    "sources": [
      "https://www.worldlabs.ai/blog/marble-world-model"
    ],
    "imported_on": "2026-07-28",
    "title": "World Labs 发布 Marble",
    "summary": "从多种输入生成 3D 世界，并支持交互编辑、扩展和组合。"
  },
  {
    "id": "gwm-1-2025-release",
    "project_id": "gwm-1",
    "date": "2025-12-11",
    "event": "Runway announced GWM-1 as three separately post-trained world-model variants.",
    "sources": [
      "https://runway.com/research/introducing-runway-gwm-1"
    ],
    "imported_on": "2026-07-27",
    "title": "Runway 介绍 GWM-1 家族",
    "summary": "基于 Gen-4.5、逐帧生成且可通过动作控制的实时自回归模型。"
  },
  {
    "id": "ctrl-world-iclr-2026",
    "project_id": "ctrl-world",
    "date": "2026",
    "event": "Ctrl-World studied policy-in-the-loop rollouts in an action-conditioned robot world model at ICLR 2026.",
    "sources": [
      "https://ctrl-world.github.io/"
    ],
    "imported_on": "2026-10-01",
    "title": "Ctrl-World 研究机器人策略在生成环境中的交互",
    "summary": "Ctrl-World 根据初始画面和逐帧动作条件生成多视角后续画面，用于让机器人策略在学得的环境中交互和评估。"
  },
  {
    "id": "leworldmodel-2026-research",
    "project_id": "leworldmodel",
    "date": "2026",
    "event": "LeWorldModel presented end-to-end predictive world-model research in 2026.",
    "sources": [
      "https://arxiv.org/abs/2603.19312"
    ],
    "imported_on": "2026-10-01",
    "title": "LeWorldModel 研究从像素端到端学习预测表征",
    "summary": "LeWorldModel 研究稳定的端到端 JEPA 训练，在潜在空间根据观测与动作预测后续表征，并用于目标条件规划。"
  },
  {
    "id": "waymo-world-model-2026-release",
    "project_id": "waymo-world-model",
    "date": "2026-02-06",
    "event": "Waymo introduced a Genie 3-derived driving world model that emits camera and lidar simulation.",
    "sources": [
      "https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-frontier-for-autonomous-driving-simulation/"
    ],
    "imported_on": "2026-07-27",
    "title": "Waymo 介绍驾驶世界模型",
    "summary": "基于 Genie 3 的 Waymo 世界模型。"
  },
  {
    "id": "cosmos-3-2026-release",
    "project_id": "cosmos-3",
    "date": "2026-06-01",
    "event": "NVIDIA launched Cosmos 3 as an open-weight physical-AI model.",
    "sources": [
      "https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Launches-Cosmos-3-the-Open-Frontier-Foundation-Model-for-Physical-AI/default.aspx",
      "https://arxiv.org/abs/2606.02800v4"
    ],
    "imported_on": "2026-08-06",
    "title": "NVIDIA 发布 Cosmos 3",
    "summary": "NVIDIA 发布面向物理 AI 的 Cosmos 3，介绍机器人、自动驾驶与视觉智能体等应用方向。"
  },
  {
    "id": "gaia-4-2026-release",
    "project_id": "gaia-4",
    "date": "2026-08-03",
    "event": "Wayve introduced GAIA-4 as the world model at the core of its closed-loop Simulation 2.0 system.",
    "sources": [
      "https://wayve.ai/thinking/gaia-4/"
    ],
    "imported_on": "2026-08-05",
    "title": "Wayve 介绍 GAIA-4",
    "summary": "让 AI Driver 参与闭环模拟的 GAIA 世界模型。"
  },
  {
    "id": "alayaworld-v1-1-report",
    "project_id": "alayaworld-v1-1",
    "date": "2026-08-13",
    "event": "The AlayaWorld v1.1 technical report described streaming spatial memory and revised video conditioning.",
    "sources": [
      "https://arxiv.org/html/2608.13492v1"
    ],
    "imported_on": "2026-10-10",
    "title": "AlayaWorld v1.1 公开技术报告",
    "summary": "报告说明以流式三维点缓存提供空间记忆和视角条件，修订条件编码与视频生成的衔接；内部三维记忆服务于视频预测。"
  },
  {
    "id": "alaya-evoke-v2-paper",
    "project_id": "alaya-evoke",
    "date": "2026-08-18",
    "event": "The Alaya-EVOKE paper reached v2, describing external world-state memory for interactive generation.",
    "sources": [
      "https://arxiv.org/html/2608.13546v2",
      "https://arxiv.org/abs/2608.13546"
    ],
    "imported_on": "2026-10-10",
    "title": "Alaya-EVOKE 论文更新至 v2",
    "summary": "该版本以外部世界状态库保存相机相关场景信息，取回当前视角内容用于持续生成；固定预算的记忆仍不保证长时一致。"
  },
  {
    "id": "pixverse-r2-2026-research",
    "project_id": "pixverse-r2",
    "date": "2026-08-23",
    "event": "PixVerse published the R2 technical description of continuously controlled real-time audiovisual generation.",
    "sources": [
      "https://pixverse.ai/en/blog/pixverse-r2-scaling-real-time-omni-world-models"
    ],
    "imported_on": "2026-10-10",
    "title": "PixVerse 介绍 R2 实时视听世界模型技术",
    "summary": "官方技术说明介绍以历史状态和实时多模态输入持续驱动同步音视频生成的 R2，并说明流式预测、记忆与实时加速方案。"
  },
  {
    "id": "solaris-2026-introduction",
    "project_id": "solaris",
    "date": "2026-08-31",
    "event": "Runway introduced Solaris, a frame-by-frame generative interface model, with early-access applications.",
    "sources": [
      "https://runway.com/news/research/introducing-solaris"
    ],
    "imported_on": "2026-10-10",
    "title": "Runway 介绍逐帧生成交互界面的 Solaris",
    "summary": "Runway 介绍由世界模型生成视觉界面的 Solaris，点击、拖拽等交互作为后续画面的条件，并提供早期访问申请。"
  },
  {
    "id": "marble-world-labs-atlas",
    "project_id": "atlas",
    "date": "2026-09-01",
    "event": "World Labs introduced Atlas, a multimodal world model for spatial intelligence.",
    "sources": [
      "https://www.worldlabs.ai/blog/atlas"
    ],
    "imported_on": "2026-10-01",
    "title": "World Labs 介绍 Atlas 世界模型",
    "summary": "World Labs 介绍 Atlas，以共享空间上下文生成图像、视频和三维表示，并宣布向部分合作伙伴提供早期访问。"
  },
  {
    "id": "gwm-1-runway-worlds-2",
    "project_id": "gwm-worlds-2",
    "date": "2026-09-03",
    "event": "Runway introduced GWM Worlds 2 as a research preview for interactive audiovisual world generation.",
    "sources": [
      "https://runway.com/research/introducing-gwm-worlds-2"
    ],
    "imported_on": "2026-10-10",
    "title": "GWM Worlds 2：音视频交互世界研究预览",
    "summary": "Runway 介绍后续版本 GWM Worlds 2：生成 720p、24 fps 视频及音频，响应连续相机控制与文本输入。官方同时指出快速旋转可能损伤细节，长时记忆仍不完善。"
  },
  {
    "id": "marble-world-labs-amd-agreement",
    "project_id": "marble",
    "date": "2026-09-28",
    "event": "World Labs announced that it had signed a definitive agreement to join AMD; closing was expected by the end of 2026, subject to regulatory approvals and customary closing conditions.",
    "sources": [
      "https://www.worldlabs.ai/blog/amd-announcement"
    ],
    "imported_on": "2026-10-05",
    "title": "World Labs 宣布签署加入 AMD 的协议",
    "summary": "World Labs 宣布已与 AMD 签署最终协议，拟加入 AMD。公告预计交易于 2026 年底完成，仍须获得监管批准并满足其他惯常交割条件。"
  },
  {
    "id": "v2x-wam-v2-paper",
    "project_id": "v2x-wam",
    "date": "2026-09-30",
    "event": "The V2X-WAM paper reached v2, describing cooperative observations and trajectory-conditioned future occupancy and flow.",
    "sources": [
      "https://arxiv.org/html/2609.37098v2"
    ],
    "imported_on": "2026-10-10",
    "title": "V2X-WAM 论文更新至 v2",
    "summary": "该版本结合车辆与路侧观测生成候选轨迹，并以选择的轨迹和动作序列预测未来占据与动态流。"
  },
  {
    "id": "uniwam-v1-paper",
    "project_id": "uniwam",
    "date": "2026-10-01",
    "event": "The UniWAM v1 paper introduced joint semantic understanding, visual prediction, and action generation.",
    "sources": [
      "https://arxiv.org/html/2610.02054v1",
      "https://arxiv.org/abs/2610.02054"
    ],
    "imported_on": "2026-10-10",
    "title": "UniWAM 提出联合理解、视觉预测与动作生成",
    "summary": "论文介绍将物理推理器、世界生成器与动作预测器放入统一架构，共同学习语言、未来视觉和动作。"
  },
  {
    "id": "deltaworld-v1-paper",
    "project_id": "deltaworld",
    "date": "2026-10-02",
    "event": "The DeltaWorld v1 paper introduced action-conditioned latent increment prediction for robot manipulation videos.",
    "sources": [
      "https://arxiv.org/html/2610.02691v1",
      "https://arxiv.org/abs/2610.02691"
    ],
    "imported_on": "2026-10-10",
    "title": "DeltaWorld 提出动作条件的状态增量预测",
    "summary": "论文先预测机器人动作引起的潜在状态增量，再更新状态并解码未来画面，以关注操作中局部物体的变化。"
  },
  {
    "id": "h-jepa-v1-paper",
    "project_id": "h-jepa",
    "date": "2026-10-05",
    "event": "The H-JEPA v1 paper introduced a hierarchy of action-conditioned latent prediction and planning.",
    "sources": [
      "https://arxiv.org/html/2610.06805v1"
    ],
    "imported_on": "2026-10-10",
    "title": "H-JEPA 提出分层未来表征预测",
    "summary": "论文介绍在不同时间尺度学习动作条件潜在状态；上层预测为下层规划提供子目标，实体机器人闭环控制仍待验证。"
  },
  {
    "id": "token-world-v2-paper",
    "project_id": "token-world",
    "date": "2026-10-05",
    "event": "The Token-World paper reached v2, describing robot world simulation in policy-facing VLM token space.",
    "sources": [
      "https://arxiv.org/html/2610.00575v2"
    ],
    "imported_on": "2026-10-10",
    "title": "Token-World 论文更新至 v2",
    "summary": "该版本在策略使用的 VLM token 空间中，按机器人动作预测后续状态；RGB 重建不参与主要预测过程，代码仍是发布承诺。"
  },
  {
    "id": "ppwm-v1-paper",
    "project_id": "ppwm",
    "date": "2026-10-06",
    "event": "The PPWM v1 paper proposed causal parallel prediction of finite-horizon latent trajectories for planning.",
    "sources": [
      "https://arxiv.org/html/2610.08627v1"
    ],
    "imported_on": "2026-10-10",
    "title": "PPWM 提出有限跨度的并行未来预测",
    "summary": "论文让未来表征之间进行因果交互并并行预测，用候选动作序列辅助规划；预测在单段内并行，超出跨度后仍需分段递推。"
  },
  {
    "id": "long-wam-v1-paper",
    "project_id": "long-wam",
    "date": "2026-10-07",
    "event": "The Long-WAM v1 paper described longer visual context for future prediction and robot action generation.",
    "sources": [
      "https://arxiv.org/html/2610.10528v1"
    ],
    "imported_on": "2026-10-10",
    "title": "Long-WAM 公开长视觉历史的世界动作模型研究",
    "summary": "论文利用较长视觉历史预测未来视频表征，再生成机器人动作块，并比较逆动力学与视频动作联合去噪两种路径。"
  },
  {
    "id": "dreamtrue-v1-paper",
    "project_id": "dreamtrue",
    "date": "2026-10-08",
    "event": "The DreamTrue v1 paper described multi-view action-conditioned robot video prediction and counterfactual post-training.",
    "sources": [
      "https://arxiv.org/html/2610.12468v1"
    ],
    "imported_on": "2026-10-10",
    "title": "DreamTrue 公开动作条件机器人视频预测研究",
    "summary": "论文以机器人动作轨迹为条件预测多视角未来视频，通过几何校准和反事实后训练改善动作遵循；遮挡下仍可能出现物理错误。"
  },
  {
    "id": "odyssey-3-2026-release",
    "project_id": "odyssey-3",
    "date": "2026-10-08",
    "event": "Odyssey introduced Odyssey-3 and its Flash research preview for interactive visual environments.",
    "sources": [
      "https://odyssey.systems/meet-odyssey-3"
    ],
    "imported_on": "2026-10-10",
    "title": "Odyssey 发布 Odyssey-3 研究预览",
    "summary": "官方介绍用已有观察与当前动作持续预测视觉环境的 Odyssey-3，并提供 Flash 的研究预览，支持导航与独立相机移动。"
  },
  {
    "id": "planwam-v1-paper",
    "project_id": "planwam",
    "date": "2026-10-08",
    "event": "The PlanWAM v1 paper described history-conditioned future representations for autonomous-driving planning.",
    "sources": [
      "https://arxiv.org/html/2610.11382v1"
    ],
    "imported_on": "2026-10-10",
    "title": "PlanWAM 公开面向驾驶规划的未来表征研究",
    "summary": "论文从历史观测预测有助于驾驶规划的未来潜在表征，再生成并选择轨迹。真实未来仅用于训练，推理阶段只读取已有历史。"
  }
]