{
  "schema_version": 1,
  "projects": {
    "world-models-2018": {
      "route": {
        "value": "latent",
        "quote": "The MDN outputs the parameters of a mixture of Gaussian distribution used to sample a prediction of the next latent vector",
        "source_index": 1,
        "source_url": "https://worldmodels.github.io/",
        "observed_on": "2026-09-21"
      },
      "one_liner": {
        "value": "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。",
        "quote": "We can even train our agent entirely inside of its own dream environment generated by its world model, and transfer this policy back into the actual environment.",
        "source_index": 1,
        "source_url": "https://worldmodels.github.io/",
        "observed_on": "2026-09-21"
      },
      "output": {
        "value": "用于采样下一潜在向量预测的高斯混合分布参数。",
        "quote": "The MDN outputs the parameters of a mixture of Gaussian distribution used to sample a prediction of the next latent vector",
        "source_index": 1,
        "source_url": "https://worldmodels.github.io/",
        "observed_on": "2026-09-21"
      },
      "input": {
        "value": "记忆模型 M 接收当前潜在向量和动作。",
        "quote": "M will then take the current $z_{t}$z​t​​ and action $a_{t}$a​t​​ as an input",
        "source_index": 1,
        "source_url": "https://worldmodels.github.io/",
        "observed_on": "2026-09-26"
      },
      "action_conditioned": {
        "value": true,
        "quote": "M will then take the current $z_{t}$z​t​​ and action $a_{t}$a​t​​ as an input",
        "source_index": 1,
        "source_url": "https://worldmodels.github.io/",
        "observed_on": "2026-09-26"
      }
    },
    "sora-world-simulators": {
      "route": {
        "value": "pixel",
        "quote": "Our largest model, Sora, is capable of generating a minute of high fidelity video.",
        "source_index": 1,
        "source_url": "https://openai.com/index/video-generation-models-as-world-simulators/",
        "observed_on": "2026-09-21"
      },
      "one_liner": {
        "value": "OpenAI 在这篇技术报告中介绍可生成最长一分钟视频的 Sora。",
        "quote": "it can generate videos and images spanning diverse durations, aspect ratios and resolutions, up to a full minute of high definition video.",
        "source_index": 1,
        "source_url": "https://openai.com/index/video-generation-models-as-world-simulators/",
        "observed_on": "2026-10-10",
        "snapshot_hash": "150de4c2fe3955be595ce9abce23a2d64918752dc7b7a795d183da3ef3bd67b7"
      },
      "output": {
        "value": "一分钟视频。",
        "quote": "Our largest model, Sora, is capable of generating a minute of high fidelity video.",
        "source_index": 1,
        "source_url": "https://openai.com/index/video-generation-models-as-world-simulators/",
        "observed_on": "2026-09-21"
      },
      "input": {
        "value": "图像和提示词。",
        "quote": "Sora is capable of generating videos provided an image and prompt as input.",
        "source_index": 1,
        "source_url": "https://openai.com/index/video-generation-models-as-world-simulators/",
        "observed_on": "2026-09-21"
      },
      "limit": {
        "value": "该报告指出，Sora 不能准确建模一些基础物理交互，例如玻璃破碎。",
        "quote": "Sora currently exhibits numerous limitations as a simulator. For example, it does not accurately model the physics of many basic interactions, like glass shattering.",
        "source_index": 1,
        "source_url": "https://openai.com/index/video-generation-models-as-world-simulators/",
        "observed_on": "2026-09-21"
      }
    },
    "copilot4d": {
      "route": {
        "value": "latent",
        "quote": "Use these BEV tokens to predict future BEV tokens with our world model auto-regressively, which leverages discrete diffusion.",
        "source_index": 1,
        "source_url": "https://waabi.ai/research/copilot-4d",
        "observed_on": "2026-09-21"
      },
      "one_liner": {
        "value": "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。",
        "quote": "Consequently, we propose Copilot4D, a novel world modeling approach that first tokenizes sensor observations with VQVAE, then predicts the future via discrete diffusion.",
        "source_index": 1,
        "source_url": "https://waabi.ai/research/copilot-4d",
        "observed_on": "2026-09-21"
      },
      "input": {
        "value": "BEV token 和自车动作。",
        "quote": "With those BEV tokens as well as the SDV action at t, it performs k steps of discrete diffusion to obtain the BEV tokens representing the next frame, t.",
        "source_index": 1,
        "source_url": "https://waabi.ai/research/copilot-4d",
        "observed_on": "2026-09-21"
      },
      "output": {
        "value": "由预测的 BEV token 解码得到的未来点云。",
        "quote": "Map the predicted futures from BEV tokens back to point clouds with the tokenizer decoder and differentiable neural depth rendering.",
        "source_index": 1,
        "source_url": "https://waabi.ai/research/copilot-4d",
        "observed_on": "2026-09-21"
      },
      "action_conditioned": {
        "value": true,
        "quote": "To make this possible, Copilot4D has the ability to be prompted with potential future actions, and thus it can be called a world model.",
        "source_index": 1,
        "source_url": "https://waabi.ai/research/copilot-4d",
        "observed_on": "2026-09-21"
      }
    },
    "genie-3": {
      "route": {
        "value": "interactive",
        "quote": "Given a text prompt, Genie 3 can generate dynamic worlds that you can navigate in real time at 24 frames per second, retaining consistency for a few minutes at a resolution of 720p.",
        "source_index": 1,
        "source_url": "https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/",
        "observed_on": "2026-09-21"
      },
      "one_liner": {
        "value": "根据文本提示生成可实时导航的动态世界。",
        "quote": "Given a text prompt, Genie 3 can generate dynamic worlds that you can navigate in real time at 24 frames per second, retaining consistency for a few minutes at a resolution of 720p.",
        "source_index": 1,
        "source_url": "https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/",
        "observed_on": "2026-09-21"
      },
      "input": {
        "value": "导航输入和文本交互。",
        "quote": "In addition to navigational inputs, Genie 3 also enables a more expressive form of text-based interaction, which we refer to as promptable world events.",
        "source_index": 1,
        "source_url": "https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/",
        "observed_on": "2026-09-21"
      },
      "output": {
        "value": "可实时导航的动态视觉世界。",
        "quote": "Given a text prompt, Genie 3 can generate dynamic worlds that you can navigate in real time at 24 frames per second, retaining consistency for a few minutes at a resolution of 720p.",
        "source_index": 1,
        "source_url": "https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/",
        "observed_on": "2026-09-21"
      },
      "action_conditioned": {
        "value": true,
        "quote": "Like any other environment, Genie 3 is not aware of the agent’s goal, instead it simulates the future based on the agent's actions.",
        "source_index": 1,
        "source_url": "https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/",
        "observed_on": "2026-09-21"
      },
      "public": {
        "value": "limited_access",
        "quote": "we are announcing Genie 3 as a limited research preview, providing early access to a small cohort of academics and creators.",
        "source_index": 1,
        "source_url": "https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/",
        "observed_on": "2026-09-21"
      },
      "limit": {
        "value": "该发布文称，连续交互只能维持几分钟。",
        "quote": "Limited interaction duration. The model can currently support a few minutes of continuous interaction, rather than extended hours.",
        "source_index": 1,
        "source_url": "https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/",
        "observed_on": "2026-09-21"
      }
    },
    "vjepa-2-ac": {
      "route": {
        "value": "latent",
        "quote": "post-training a latent action-conditioned world model, V-JEPA 2-AC, using less than 62 hours of unlabeled robot videos from the Droid dataset.",
        "source_index": 2,
        "source_url": "https://arxiv.org/abs/2506.09985v1",
        "observed_on": "2026-09-21"
      },
      "action_conditioned": {
        "value": true,
        "quote": "post-training a latent action-conditioned world model, V-JEPA 2-AC, using less than 62 hours of unlabeled robot videos from the Droid dataset.",
        "source_index": 2,
        "source_url": "https://arxiv.org/abs/2506.09985v1",
        "observed_on": "2026-09-21"
      },
      "one_liner": {
        "value": "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。",
        "quote": "After pretraining, we freeze the video encoder and learn a new action-conditioned predictor, V-JEPA 2-AC, on top of the learned representation.",
        "source_index": 3,
        "source_url": "https://arxiv.org/pdf/2506.09985v1",
        "observed_on": "2026-09-21"
      },
      "confused_with": {
        "value": [
          "V-JEPA 2 基础视频表征模型"
        ],
        "quote": "After pretraining, we freeze the video encoder and learn a new action-conditioned predictor, V-JEPA 2-AC, on top of the learned representation.",
        "source_index": 3,
        "source_url": "https://arxiv.org/pdf/2506.09985v1",
        "observed_on": "2026-09-21"
      },
      "input": {
        "value": "过去的视频帧、动作和末端执行器状态。",
        "quote": "predicting the representations of future video frames conditioned on past video frames, actions, and end-effector states.",
        "source_index": 3,
        "source_url": "https://arxiv.org/pdf/2506.09985v1",
        "observed_on": "2026-09-21"
      },
      "output": {
        "value": "以控制动作和本体感知观测为条件，预测未来视频观测的表征。",
        "quote": "To achieve this, we train V-JEPA 2-AC, an autoregressive model that predicts representations of future video observations conditioned on control actions and proprioceptive observations.",
        "source_index": 3,
        "source_url": "https://arxiv.org/pdf/2506.09985v1",
        "observed_on": "2026-09-21"
      },
      "limit": {
        "value": "自回归预测会累积误差，限制长时程规划的可靠性。",
        "quote": "Long horizon planning with world models is limited by a number of factors.\nFirst, autoregressive prediction suffers from error accumulation:",
        "source_index": 3,
        "source_url": "https://arxiv.org/pdf/2506.09985v1",
        "observed_on": "2026-09-21"
      },
      "public": {
        "value": "open_weights",
        "quote": "[checkpoint](https://dl.fbaipublicfiles.com/vjepa2/vjepa2-ac-vitg.pt)",
        "source_index": 4,
        "source_url": "https://github.com/facebookresearch/vjepa2",
        "observed_on": "2026-09-26"
      }
    },
    "marble": {
      "route": {
        "value": "interactive",
        "quote": "Marble also lets you interactively edit, expand, and combine worlds.",
        "source_index": 1,
        "source_url": "https://www.worldlabs.ai/blog/marble-world-model",
        "observed_on": "2026-09-21"
      },
      "one_liner": {
        "value": "从多种输入生成 3D 世界，并支持交互编辑、扩展和组合。",
        "quote": "Marble can create 3D worlds from text, images, video, or coarse 3D layouts; Marble also lets you interactively edit, expand, and combine worlds.",
        "source_index": 1,
        "source_url": "https://www.worldlabs.ai/blog/marble-world-model",
        "observed_on": "2026-09-21"
      },
      "input": {
        "value": "文本、图像、视频或粗略 3D 布局。",
        "quote": "Marble can create 3D worlds from text, images, video, or coarse 3D layouts;",
        "source_index": 1,
        "source_url": "https://www.worldlabs.ai/blog/marble-world-model",
        "observed_on": "2026-09-21"
      },
      "output": {
        "value": "可导出为 Gaussian splats、网格或视频的 3D 世界。",
        "quote": "Once generated, 3D worlds can be exported as Gaussian splats, meshes, or videos.",
        "source_index": 1,
        "source_url": "https://www.worldlabs.ai/blog/marble-world-model",
        "observed_on": "2026-09-21"
      },
      "public": {
        "value": "public",
        "quote": "Today we are making Marble, a first-in-class generative multimodal world model, generally available for anyone to use.",
        "source_index": 1,
        "source_url": "https://www.worldlabs.ai/blog/marble-world-model",
        "observed_on": "2026-09-21"
      },
      "limit": {
        "value": "碰撞网格为低保真网格，用于粗略物理模拟。",
        "quote": "Marble can generate both collider meshes, which are low-fidelity meshes intended for coarse physics simulation;",
        "source_index": 1,
        "source_url": "https://www.worldlabs.ai/blog/marble-world-model",
        "observed_on": "2026-09-21"
      }
    },
    "gwm-1": {
      "one_liner": {
        "value": "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。",
        "quote": "GWM-1 is an autoregressive model built on top of Gen-4.5. It generates frame by frame, runs in real time, and can be controlled interactively with actions—camera pose, robot commands, audio.",
        "source_index": 1,
        "source_url": "https://runway.com/research/introducing-runway-gwm-1",
        "observed_on": "2026-09-21"
      },
      "route": {
        "value": "interactive",
        "quote": "It generates frame by frame, runs in real time, and can be controlled interactively with actions—camera pose, robot commands, audio.",
        "source_index": 1,
        "source_url": "https://runway.com/research/introducing-runway-gwm-1",
        "observed_on": "2026-09-21"
      },
      "input": {
        "value": "家族所述动作控制包括相机位姿、机器人指令和音频。",
        "quote": "It generates frame by frame, runs in real time, and can be controlled interactively with actions—camera pose, robot commands, audio.",
        "source_index": 1,
        "source_url": "https://runway.com/research/introducing-runway-gwm-1",
        "observed_on": "2026-09-21"
      },
      "output": {
        "value": "逐帧生成的视觉序列。",
        "quote": "It generates frame by frame, runs in real time, and can be controlled interactively with actions—camera pose, robot commands, audio.",
        "source_index": 1,
        "source_url": "https://runway.com/research/introducing-runway-gwm-1",
        "observed_on": "2026-09-21"
      },
      "action_conditioned": {
        "value": true,
        "quote": "It generates frame by frame, runs in real time, and can be controlled interactively with actions—camera pose, robot commands, audio.",
        "source_index": 1,
        "source_url": "https://runway.com/research/introducing-runway-gwm-1",
        "observed_on": "2026-09-21"
      },
      "confused_with": {
        "value": [
          "单一、统一的通用模型"
        ],
        "quote": "GWM Worlds for explorable environments, GWM Avatars for conversational characters, and GWM Robotics for robotic manipulation. Today, these are separate post-trained models.",
        "source_index": 1,
        "source_url": "https://runway.com/research/introducing-runway-gwm-1",
        "observed_on": "2026-09-21"
      }
    },
    "waymo-world-model": {
      "one_liner": {
        "value": "基于 Genie 3 的 Waymo 世界模型。",
        "quote": "The Waymo World Model is built upon Genie 3",
        "source_index": 1,
        "source_url": "https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-frontier-for-autonomous-driving-simulation/",
        "observed_on": "2026-09-21"
      },
      "route": {
        "value": "interactive",
        "quote": "Driving action control allows us to have a responsive simulator that adheres to specific driving inputs.",
        "source_index": 1,
        "source_url": "https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-frontier-for-autonomous-driving-simulation/",
        "observed_on": "2026-09-21"
      },
      "action_conditioned": {
        "value": true,
        "quote": "Driving action control allows us to have a responsive simulator that adheres to specific driving inputs.",
        "source_index": 1,
        "source_url": "https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-frontier-for-autonomous-driving-simulation/",
        "observed_on": "2026-09-21"
      },
      "input": {
        "value": "语言提示、驾驶输入和场景布局。",
        "quote": "The model’s architecture offers high controllability, allowing our engineers to modify simulations with simple language prompts, driving inputs, and scene layouts.",
        "source_index": 1,
        "source_url": "https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-frontier-for-autonomous-driving-simulation/",
        "observed_on": "2026-09-21"
      },
      "output": {
        "value": "包含相机与激光雷达数据的多传感器模拟输出。",
        "quote": "Notably, the Waymo World Model generates high-fidelity, multi-sensor outputs that include both camera and lidar data.",
        "source_index": 1,
        "source_url": "https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-frontier-for-autonomous-driving-simulation/",
        "observed_on": "2026-09-21"
      },
      "limit": {
        "value": "模拟时间越长，计算和维持稳定质量越困难。",
        "quote": "That’s harder to do because the longer the simulation, the tougher it is to compute and maintain stable quality.",
        "source_index": 1,
        "source_url": "https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-frontier-for-autonomous-driving-simulation/",
        "observed_on": "2026-09-26"
      }
    },
    "cosmos-3": {
      "route": {
        "value": "physics",
        "quote": "NVIDIA today launched NVIDIA Cosmos™ 3, an open world foundation model for physical AI",
        "source_index": 1,
        "source_url": "https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Launches-Cosmos-3-the-Open-Frontier-Foundation-Model-for-Physical-AI/default.aspx",
        "observed_on": "2026-09-21"
      },
      "one_liner": {
        "value": "联合处理和生成语言、图像、视频、音频及动作序列的世界模型家族。",
        "quote": "We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences",
        "source_index": 2,
        "source_url": "https://arxiv.org/abs/2606.02800v4",
        "observed_on": "2026-09-21"
      },
      "input": {
        "value": "语言、图像、视频、音频和动作序列。",
        "quote": "We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences",
        "source_index": 2,
        "source_url": "https://arxiv.org/abs/2606.02800v4",
        "observed_on": "2026-09-21"
      },
      "output": {
        "value": "语言、图像、视频、音频和动作序列。",
        "quote": "We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences",
        "source_index": 2,
        "source_url": "https://arxiv.org/abs/2606.02800v4",
        "observed_on": "2026-09-21"
      },
      "public": {
        "value": "open_weights",
        "quote": "Cosmos 3 Edge, Cosmos 3 Nano and Cosmos 3 Super are available now on_[_Hugging Face_](https://huggingface.co/collections/nvidia/cosmos3)",
        "source_index": 3,
        "source_url": "https://blogs.nvidia.com/blog/siggraph-news-2026/",
        "observed_on": "2026-09-26"
      },
      "action_conditioned": {
        "value": true,
        "quote": "Generate infinite plausible futures from text, image, video, ambient sound and action input.",
        "source_index": 4,
        "source_url": "https://www.nvidia.com/en-us/ai/cosmos/",
        "observed_on": "2026-09-26"
      }
    },
    "gaia-4": {
      "one_liner": {
        "value": "让 AI Driver 参与闭环模拟的 GAIA 世界模型。",
        "quote": "Now with GAIA-4, we can close the loop, turning a recorded scene into the starting point for a new simulated run with the AI Driver back in the loop:",
        "source_index": 1,
        "source_url": "https://wayve.ai/thinking/gaia-4/",
        "observed_on": "2026-09-21"
      },
      "route": {
        "value": "interactive",
        "quote": "the AI Driver’s actions drive the simulation forward — its position, its viewpoint, and the sensor stream it receives",
        "source_index": 1,
        "source_url": "https://wayve.ai/thinking/gaia-4/",
        "observed_on": "2026-09-21"
      },
      "action_conditioned": {
        "value": true,
        "quote": "the AI Driver’s actions drive the simulation forward — its position, its viewpoint, and the sensor stream it receives",
        "source_index": 1,
        "source_url": "https://wayve.ai/thinking/gaia-4/",
        "observed_on": "2026-09-21"
      },
      "input": {
        "value": "记录的传感器数据。",
        "quote": "It generates the scene directly from recorded sensor data.",
        "source_index": 1,
        "source_url": "https://wayve.ai/thinking/gaia-4/",
        "observed_on": "2026-09-21"
      },
      "output": {
        "value": "相机与雷达的多模态模拟数据。",
        "quote": "GAIA-4 extends this capability with multimodal generation, bringing radar into the simulation alongside cameras, an industry-first for autonomous driving simulators.",
        "source_index": 1,
        "source_url": "https://wayve.ai/thinking/gaia-4/",
        "observed_on": "2026-09-21"
      },
      "limit": {
        "value": "World-on-rails 是可选回放约束，并非模型整体限制；不能把该模式等同于完全交互的交通模拟。",
        "quote": "World-on-rails is a capability we can switch on, not a limitation of the world model.",
        "source_index": 1,
        "source_url": "https://wayve.ai/thinking/gaia-4/",
        "observed_on": "2026-09-26"
      },
      "confused_with": {
        "value": [
          "world-on-rails 模式下完全响应自车的交通模拟"
        ],
        "quote": "World-on-rails means every road user follows its logged trajectory, producing a conservative replay used for validation.",
        "source_index": 1,
        "source_url": "https://wayve.ai/thinking/gaia-4/",
        "observed_on": "2026-09-21"
      }
    },
    "dreamerv3": {
      "route": {
        "value": "latent",
        "quote": "The world model learns compact representations of sensory inputs through autoencoding24 and enables planning by predicting future representations and rewards for potential actions.",
        "source_index": 1,
        "source_url": "https://www.nature.com/articles/s41586-025-08744-2",
        "observed_on": "2026-09-30",
        "snapshot_hash": "17cc010c487fa21ea829865c7f4eab38010046355adf9737cc5d34b5b7270a2c"
      },
      "one_liner": {
        "value": "学习环境模型，通过想象未来情景改进智能体行为。",
        "quote": "Dreamer learns a model of the environment and improves its behaviour by imagining future scenarios.",
        "source_index": 1,
        "source_url": "https://www.nature.com/articles/s41586-025-08744-2",
        "observed_on": "2026-09-30",
        "snapshot_hash": "17cc010c487fa21ea829865c7f4eab38010046355adf9737cc5d34b5b7270a2c"
      },
      "input": {
        "value": "感官观测及动作。",
        "quote": "The world model encodes sensory inputs xt using the encoder (enc) into discrete representations zt that are predicted by a sequence model with recurrent state ht given actions at.",
        "source_index": 1,
        "source_url": "https://www.nature.com/articles/s41586-025-08744-2",
        "observed_on": "2026-09-30",
        "snapshot_hash": "17cc010c487fa21ea829865c7f4eab38010046355adf9737cc5d34b5b7270a2c"
      },
      "output": {
        "value": "未来内部表征和奖励。",
        "quote": "enables planning by predicting future representations and rewards for potential actions.",
        "source_index": 1,
        "source_url": "https://www.nature.com/articles/s41586-025-08744-2",
        "observed_on": "2026-09-30",
        "snapshot_hash": "17cc010c487fa21ea829865c7f4eab38010046355adf9737cc5d34b5b7270a2c"
      },
      "action_conditioned": {
        "value": true,
        "quote": "the world model predicts the outcomes of potential actions",
        "source_index": 1,
        "source_url": "https://www.nature.com/articles/s41586-025-08744-2",
        "observed_on": "2026-09-30",
        "snapshot_hash": "17cc010c487fa21ea829865c7f4eab38010046355adf9737cc5d34b5b7270a2c"
      },
      "open_source": {
        "value": true,
        "quote": "Source code for Dreamer and all presented experiments is available under an open-source license at https://github.com/danijar/dreamerv3.",
        "source_index": 1,
        "source_url": "https://www.nature.com/articles/s41586-025-08744-2",
        "observed_on": "2026-09-30",
        "snapshot_hash": "17cc010c487fa21ea829865c7f4eab38010046355adf9737cc5d34b5b7270a2c"
      }
    },
    "dino-wm": {
      "route": {
        "value": "latent",
        "quote": "DINO-WM models the world dynamics on compact embeddings of the world, rather than the raw observations themselves.",
        "source_index": 1,
        "source_url": "https://dino-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "58f6a84f1b1ab26f220bdb14cf54fab67e3cec8a17d5ace43812affb7cc452b9"
      },
      "one_liner": {
        "value": "将目标特征作为预测目标，通过优化动作序列进行视觉目标规划。",
        "quote": "This allows DINO-WM to achieve observational goals through action sequence optimization, facilitating task-agnostic planning by treating goal features as prediction targets.",
        "source_index": 1,
        "source_url": "https://dino-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "58f6a84f1b1ab26f220bdb14cf54fab67e3cec8a17d5ace43812affb7cc452b9"
      },
      "input": {
        "value": "视觉表征和动作。",
        "quote": "Given these visual embeddings and actions, DINO-WM uses the ViT architecture to predict future embeddings.",
        "source_index": 1,
        "source_url": "https://dino-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "58f6a84f1b1ab26f220bdb14cf54fab67e3cec8a17d5ace43812affb7cc452b9"
      },
      "output": {
        "value": "未来视觉 patch 特征。",
        "quote": "DINO-WM leverages spatial patch features pre-trained with DINOv2, enabling it to learn from offline behavioral trajectories by predicting future patch features.",
        "source_index": 1,
        "source_url": "https://dino-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "58f6a84f1b1ab26f220bdb14cf54fab67e3cec8a17d5ace43812affb7cc452b9"
      },
      "action_conditioned": {
        "value": true,
        "quote": "Given these visual embeddings and actions, DINO-WM uses the ViT architecture to predict future embeddings.",
        "source_index": 1,
        "source_url": "https://dino-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "58f6a84f1b1ab26f220bdb14cf54fab67e3cec8a17d5ace43812affb7cc452b9"
      },
      "public": {
        "value": "open_weights",
        "quote": "We have uploaded our trained world model checkpoints for PointMaze, PushT, and Wall here under checkpoints.",
        "source_index": 3,
        "source_url": "https://github.com/gaoyuezhou/dino_wm",
        "observed_on": "2026-09-30",
        "snapshot_hash": "7e98b3c3990ee5c29e7efbd0dd7b26f67145091a525ecc6ea0391d133ac16d21"
      }
    },
    "daydreamer": {
      "route": {
        "value": "latent",
        "quote": "A recurrent neural network predicts the sequence of these representations given actions.",
        "source_index": 3,
        "source_url": "https://github.com/danijar/daydreamer",
        "observed_on": "2026-09-30",
        "snapshot_hash": "ca3188c90581209658334c9dedc7a81bdcfca47112a49cff71516442b73ae55c"
      },
      "one_liner": {
        "value": "在实体机器人上学习世界模型和行为策略，不依赖模拟器。",
        "quote": "DayDreamer learns a world model and an actor critic behavior to train robots from small amounts of experience in the real world, without using simulators.",
        "source_index": 3,
        "source_url": "https://github.com/danijar/daydreamer",
        "observed_on": "2026-09-30",
        "snapshot_hash": "ca3188c90581209658334c9dedc7a81bdcfca47112a49cff71516442b73ae55c"
      },
      "input": {
        "value": "本体感知和视觉观测。",
        "quote": "To learn from proprioceptive and visual inputs alike, the world model fuses the sensory inputs of the same time step together into a compact discrete representation.",
        "source_index": 3,
        "source_url": "https://github.com/danijar/daydreamer",
        "observed_on": "2026-09-30",
        "snapshot_hash": "ca3188c90581209658334c9dedc7a81bdcfca47112a49cff71516442b73ae55c"
      },
      "output": {
        "value": "动作条件下预测的内部表征序列。",
        "quote": "A recurrent neural network predicts the sequence of these representations given actions.",
        "source_index": 3,
        "source_url": "https://github.com/danijar/daydreamer",
        "observed_on": "2026-09-30",
        "snapshot_hash": "ca3188c90581209658334c9dedc7a81bdcfca47112a49cff71516442b73ae55c"
      },
      "action_conditioned": {
        "value": true,
        "quote": "A recurrent neural network predicts the sequence of these representations given actions.",
        "source_index": 3,
        "source_url": "https://github.com/danijar/daydreamer",
        "observed_on": "2026-09-30",
        "snapshot_hash": "ca3188c90581209658334c9dedc7a81bdcfca47112a49cff71516442b73ae55c"
      },
      "limit": {
        "value": "长时间在硬件上学习会造成磨损，可能需要人工干预或维修。",
        "quote": "learning on hardware over many hours creates wear on robots that may require human intervention or repair.",
        "source_index": 4,
        "source_url": "https://arxiv.org/pdf/2206.14176v1",
        "observed_on": "2026-09-27",
        "snapshot_hash": "17976ad6089491aae16799949b3fe28421a3d81e696bb4e485ece5aba85a6cef"
      }
    },
    "diamond-atari": {
      "route": {
        "value": "pixel",
        "quote": "We train a diffusion model to predict the next frame of the game.",
        "source_index": 1,
        "source_url": "https://diamond-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "60a38ddfa17cce015d92cae27684cf6d0b8fd97ab661ec718d3023b59e9f1ea8"
      },
      "one_liner": {
        "value": "通过扩散世界模型预测游戏下一帧，并在模型中训练强化学习智能体。",
        "quote": "DIAMOND (DIffusion As a Model Of eNvironment Dreams) is a reinforcement learning agent trained entirely in a diffusion world model.",
        "source_index": 2,
        "source_url": "https://github.com/eloialonso/diamond",
        "observed_on": "2026-09-30",
        "snapshot_hash": "c721f0ea77f3ee211e47f2758aa495501c642ef09a8cbaa72fbe90e0584a80ed"
      },
      "input": {
        "value": "历史游戏帧和智能体动作。",
        "quote": "The diffusion model takes into account the agent’s action and the previous frames to simulate the environment response.",
        "source_index": 1,
        "source_url": "https://diamond-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "60a38ddfa17cce015d92cae27684cf6d0b8fd97ab661ec718d3023b59e9f1ea8"
      },
      "output": {
        "value": "下一游戏画面帧。",
        "quote": "We train a diffusion model to predict the next frame of the game.",
        "source_index": 1,
        "source_url": "https://diamond-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "60a38ddfa17cce015d92cae27684cf6d0b8fd97ab661ec718d3023b59e9f1ea8"
      },
      "action_conditioned": {
        "value": true,
        "quote": "The diffusion world model takes into account the agent's action and previous frames to generate the next frame.",
        "source_index": 1,
        "source_url": "https://diamond-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "60a38ddfa17cce015d92cae27684cf6d0b8fd97ab661ec718d3023b59e9f1ea8"
      },
      "public": {
        "value": "open_weights",
        "quote": "Quick install to try our pretrained world models using miniconda:",
        "source_index": 2,
        "source_url": "https://github.com/eloialonso/diamond",
        "observed_on": "2026-09-30",
        "snapshot_hash": "c721f0ea77f3ee211e47f2758aa495501c642ef09a8cbaa72fbe90e0584a80ed"
      }
    },
    "ctrl-world": {
      "route": {
        "value": "interactive",
        "quote": "Ctrl-World is designed for policy-in-the-loop rollouts with generalist robot policies.",
        "source_index": 1,
        "source_url": "https://ctrl-world.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "0903c41d26afea25010b8be81e70df6d2661be131f21c5f44e3c51904ea00dfc"
      },
      "one_liner": {
        "value": "可与 VLA 策略共同在想象空间中滚动生成的动作条件世界模型。",
        "quote": "Ctrl-World is an action-conditioned world model compatible with modern VLA policies and enables policy-in-the-loop rollouts entirely in imagination",
        "source_index": 2,
        "source_url": "https://github.com/Robert-gyj/Ctrl-World",
        "observed_on": "2026-09-30",
        "snapshot_hash": "ee35ca515a3a0730a216d7b2b121435181c842bd52d98bc28cacefa684e39fac"
      },
      "input": {
        "value": "初始画面和动作序列。",
        "quote": "All videos are generated by passing in the initial frame and a different sequences of actions as input.",
        "source_index": 1,
        "source_url": "https://ctrl-world.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "0903c41d26afea25010b8be81e70df6d2661be131f21c5f44e3c51904ea00dfc"
      },
      "output": {
        "value": "联合预测的多视角视频，包括腕部视角。",
        "quote": "It generates joint multi-view predictions (including wrist views), enforces fine-grained action control via frame-level conditioning,",
        "source_index": 1,
        "source_url": "https://ctrl-world.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "0903c41d26afea25010b8be81e70df6d2661be131f21c5f44e3c51904ea00dfc"
      },
      "action_conditioned": {
        "value": true,
        "quote": "enforces fine-grained action control via frame-level conditioning",
        "source_index": 1,
        "source_url": "https://ctrl-world.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "0903c41d26afea25010b8be81e70df6d2661be131f21c5f44e3c51904ea00dfc"
      },
      "public": {
        "value": "open_weights",
        "quote": "This repo include the Ctrl-World model checkpoint trained on opensourced DROID dataset (~95k trajectories, 564 scenes).",
        "source_index": 3,
        "source_url": "https://huggingface.co/yjguo/Ctrl-World",
        "observed_on": "2026-09-30",
        "snapshot_hash": "08aa1f1548dff2d6aca560913513d73c62e79aaae92876d50560a9430380ef25"
      },
      "limit": {
        "value": "模型对物理交互的建模仍可能不精确。",
        "quote": "The model can evaluate a policy’s instruction-following capability but also can be imprecise in modeling physical interactions.",
        "source_index": 2,
        "source_url": "https://github.com/Robert-gyj/Ctrl-World",
        "observed_on": "2026-09-30",
        "snapshot_hash": "ee35ca515a3a0730a216d7b2b121435181c842bd52d98bc28cacefa684e39fac"
      }
    },
    "wham-2025": {
      "route": {
        "value": "interactive",
        "quote": "WHAM can be used in world-model mode (predicting how the game evolves), action-model mode (generating plausible controller actions), or both.",
        "source_index": 1,
        "source_url": "https://www.microsoft.com/en-us/research/project/wham/wham/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "703f3b2514bc32c931dd62e7e3e43604827d3158169e6a2860e132243c7af04b"
      },
      "one_liner": {
        "value": "联合建模游戏画面与控制器动作，可生成画面、动作或二者的序列。",
        "quote": "WHAM can be used in world-model mode (predicting how the game evolves), action-model mode (generating plausible controller actions), or both.",
        "source_index": 1,
        "source_url": "https://www.microsoft.com/en-us/research/project/wham/wham/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "703f3b2514bc32c931dd62e7e3e43604827d3158169e6a2860e132243c7af04b"
      },
      "input": {
        "value": "游戏图像和控制器动作条件。",
        "quote": "a predictive model that can be conditioned on (modified) images and/or controller actions.",
        "source_index": 2,
        "source_url": "https://www.nature.com/articles/s41586-025-08600-3",
        "observed_on": "2026-09-30",
        "snapshot_hash": "270b70d9e1dd7d69638c78db49f50dbfb59f47f8c626b87657e8f09feb4d52f3"
      },
      "output": {
        "value": "预测的游戏画面和控制器动作序列，依具体模式而定。",
        "quote": "WHAM was trained on human gameplay data to predict game visuals (‘frames’) and players’ controller actions",
        "source_index": 2,
        "source_url": "https://www.nature.com/articles/s41586-025-08600-3",
        "observed_on": "2026-09-30",
        "snapshot_hash": "270b70d9e1dd7d69638c78db49f50dbfb59f47f8c626b87657e8f09feb4d52f3"
      },
      "action_conditioned": {
        "value": true,
        "quote": "a predictive model that can be conditioned on (modified) images and/or controller actions.",
        "source_index": 2,
        "source_url": "https://www.nature.com/articles/s41586-025-08600-3",
        "observed_on": "2026-09-30",
        "snapshot_hash": "270b70d9e1dd7d69638c78db49f50dbfb59f47f8c626b87657e8f09feb4d52f3"
      },
      "public": {
        "value": "open_weights",
        "quote": "We provide model weights, sample data and the WHAM Demonstrator to support interpreting, validating and extending our results.",
        "source_index": 2,
        "source_url": "https://www.nature.com/articles/s41586-025-08600-3",
        "observed_on": "2026-09-30",
        "snapshot_hash": "270b70d9e1dd7d69638c78db49f50dbfb59f47f8c626b87657e8f09feb4d52f3"
      },
      "limit": {
        "value": "论文的游戏案例聚焦单一 3D 电子游戏。",
        "quote": "Extrapolating from our use case focusing on a single 3D video game,",
        "source_index": 2,
        "source_url": "https://www.nature.com/articles/s41586-025-08600-3",
        "observed_on": "2026-09-30",
        "snapshot_hash": "270b70d9e1dd7d69638c78db49f50dbfb59f47f8c626b87657e8f09feb4d52f3"
      }
    },
    "vista-2024": {
      "route": {
        "value": "interactive",
        "quote": "Execute multi-modal actions (steering angles, speeds, commands, trajectories, goal points).",
        "source_index": 1,
        "source_url": "https://github.com/OpenDriveLab/Vista",
        "observed_on": "2026-09-30",
        "snapshot_hash": "8c76b2af17f729a77f07b764232b6ad2206cfa2fea274a51ba6942a74e30d319"
      },
      "one_liner": {
        "value": "面向驾驶、强调通用性与多种控制条件的世界模型。",
        "quote": "Vista: A Generalizable Driving World Model with High Fidelity and Versatile Controllability",
        "source_index": 1,
        "source_url": "https://github.com/OpenDriveLab/Vista",
        "observed_on": "2026-09-30",
        "snapshot_hash": "8c76b2af17f729a77f07b764232b6ad2206cfa2fea274a51ba6942a74e30d319"
      },
      "input": {
        "value": "历史帧与高层意图，包括指令和目标点。",
        "quote": "historical frames as priors for coherent long-horizon rollouts. For action controllability, we incorporate a versatile set of controls from high-level intentions (command, goal point)",
        "source_index": 2,
        "source_url": "https://arxiv.org/abs/2405.17398",
        "observed_on": "2026-09-30",
        "snapshot_hash": "e3ac703dd07400de94c25a113bc0c9c0e8d91c2ebf6b509382b937e6fbdfa195"
      },
      "output": {
        "value": "未来驾驶场景的视觉预测。",
        "quote": "Simulated futures in a wide range of driving scenarios by Vista.",
        "source_index": 1,
        "source_url": "https://github.com/OpenDriveLab/Vista",
        "observed_on": "2026-09-30",
        "snapshot_hash": "8c76b2af17f729a77f07b764232b6ad2206cfa2fea274a51ba6942a74e30d319"
      },
      "action_conditioned": {
        "value": true,
        "quote": "Execute multi-modal actions (steering angles, speeds, commands, trajectories, goal points).",
        "source_index": 1,
        "source_url": "https://github.com/OpenDriveLab/Vista",
        "observed_on": "2026-09-30",
        "snapshot_hash": "8c76b2af17f729a77f07b764232b6ad2206cfa2fea274a51ba6942a74e30d319"
      },
      "public": {
        "value": "open_weights",
        "quote": "[2024/06/06] We released the model weights v1.0 at Hugging Face and Google Drive.",
        "source_index": 1,
        "source_url": "https://github.com/OpenDriveLab/Vista",
        "observed_on": "2026-09-30",
        "snapshot_hash": "8c76b2af17f729a77f07b764232b6ad2206cfa2fea274a51ba6942a74e30d319"
      }
    },
    "atlas": {
      "route": {
        "value": "interactive",
        "quote": "Atlas takes one or more reference images and generates new views at any camera position and angle you specify.",
        "source_index": 1,
        "source_url": "https://www.worldlabs.ai/blog/atlas",
        "observed_on": "2026-09-30",
        "snapshot_hash": "7e013594e936a5baac48e4da8ea34fa04a6164ca53a52586036197d600b3366b"
      },
      "one_liner": {
        "value": "以共享空间上下文生成多模态内容的世界模型。",
        "quote": "Atlas is a multimodal autoregressive diffusion transformer. Its inputs are grounded in 3D space to form a spatial context, and it generates multimodal outputs conditioned on its context.",
        "source_index": 1,
        "source_url": "https://www.worldlabs.ai/blog/atlas",
        "observed_on": "2026-10-01",
        "snapshot_hash": "586f37bf99344f61ec4595140a1e6279ab324cc0f7b9cf91cd4e3545ca99dda7"
      },
      "input": {
        "value": "文本、图像、相机位姿和三维深度图；视频表示为图像序列。",
        "quote": "At present it can operate on text, images, camera poses, and 3D depth maps; videos are represented as sequences of images.",
        "source_index": 1,
        "source_url": "https://www.worldlabs.ai/blog/atlas",
        "observed_on": "2026-09-30",
        "snapshot_hash": "7e013594e936a5baac48e4da8ea34fa04a6164ca53a52586036197d600b3366b"
      },
      "output": {
        "value": "图像帧与三维深度信息，可组成点云或 Gaussian splats。",
        "quote": "Atlas natively operates on both 2D image frames and 3D depth maps, enabling it to output worlds as point clouds or 3D Gaussian splats.",
        "source_index": 1,
        "source_url": "https://www.worldlabs.ai/blog/atlas",
        "observed_on": "2026-09-30",
        "snapshot_hash": "7e013594e936a5baac48e4da8ea34fa04a6164ca53a52586036197d600b3366b"
      },
      "public": {
        "value": "limited_access",
        "quote": "Atlas is entering early access with select partners.",
        "source_index": 1,
        "source_url": "https://www.worldlabs.ai/blog/atlas",
        "observed_on": "2026-09-30",
        "snapshot_hash": "7e013594e936a5baac48e4da8ea34fa04a6164ca53a52586036197d600b3366b"
      },
      "confused_with": {
        "value": [
          "当前 Marble 产品"
        ],
        "quote": "Atlas will power future versions of Marble and other products from World Labs.",
        "source_index": 1,
        "source_url": "https://www.worldlabs.ai/blog/atlas",
        "observed_on": "2026-09-30",
        "snapshot_hash": "7e013594e936a5baac48e4da8ea34fa04a6164ca53a52586036197d600b3366b"
      }
    },
    "leworldmodel": {
      "route": {
        "value": "latent",
        "quote": "The predictor models the environment dynamics in latent space by predicting the embedding of the next frame observation",
        "source_index": 1,
        "source_url": "https://le-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "def20e41b40620be230aabf7c7d96b210e112fafcec4e21f6e9c123f5254eff9"
      },
      "one_liner": {
        "value": "从像素端到端训练的 JEPA 世界模型。",
        "quote": "we introduce LeWorldModel (LeWM), the first JEPA that trains stably end-to-end from raw pixels using only two loss terms",
        "source_index": 1,
        "source_url": "https://le-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "def20e41b40620be230aabf7c7d96b210e112fafcec4e21f6e9c123f5254eff9"
      },
      "input": {
        "value": "当前图像的潜在表征和动作。",
        "quote": "given the latent embedding \\(\\vz_t\\) and an action \\(\\va_t\\).",
        "source_index": 1,
        "source_url": "https://le-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "def20e41b40620be230aabf7c7d96b210e112fafcec4e21f6e9c123f5254eff9"
      },
      "output": {
        "value": "下一帧观测的潜在表征。",
        "quote": "The predictor models the environment dynamics in latent space by predicting the embedding of the next frame observation",
        "source_index": 1,
        "source_url": "https://le-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "def20e41b40620be230aabf7c7d96b210e112fafcec4e21f6e9c123f5254eff9"
      },
      "action_conditioned": {
        "value": true,
        "quote": "given the latent embedding \\(\\vz_t\\) and an action \\(\\va_t\\).",
        "source_index": 1,
        "source_url": "https://le-wm.github.io/",
        "observed_on": "2026-09-30",
        "snapshot_hash": "def20e41b40620be230aabf7c7d96b210e112fafcec4e21f6e9c123f5254eff9"
      },
      "public": {
        "value": "open_weights",
        "quote": "Pretrained LeWM checkpoints for each environment are mirrored on the Hugging Face Hub (model repos), alongside the datasets (dataset repos) in the same collection:",
        "source_index": 2,
        "source_url": "https://github.com/lucas-maes/le-wm",
        "observed_on": "2026-09-30",
        "snapshot_hash": "c81d2d49e6b8606cde6a7c267ead0683ed034509cd804031c8b8fb9ed629aa24"
      }
    },
    "h-jepa": {
      "route": {
        "value": "latent",
        "quote": "All levels predict future latent states from past latent states and actions.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.06805v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "d9d8f2b3f06f22c151ef3c9796063bf9803b3a6a40ca388db495716efb0d4e2b"
      },
      "one_liner": {
        "value": "在各层潜在空间中，用历史状态与动作预测不同时间尺度的未来。",
        "quote": "All levels predict future latent states from past latent states and actions.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.06805v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "d9d8f2b3f06f22c151ef3c9796063bf9803b3a6a40ca388db495716efb0d4e2b"
      },
      "input": {
        "value": "过去的潜在状态与动作。",
        "quote": "All levels predict future latent states from past latent states and actions.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.06805v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "d9d8f2b3f06f22c151ef3c9796063bf9803b3a6a40ca388db495716efb0d4e2b"
      },
      "output": {
        "value": "各层的未来潜在状态。",
        "quote": "All levels predict future latent states from past latent states and actions.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.06805v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "d9d8f2b3f06f22c151ef3c9796063bf9803b3a6a40ca388db495716efb0d4e2b"
      },
      "action_conditioned": {
        "value": true,
        "quote": "All levels predict future latent states from past latent states and actions.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.06805v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "d9d8f2b3f06f22c151ef3c9796063bf9803b3a6a40ca388db495716efb0d4e2b"
      },
      "public": {
        "value": "open_weights",
        "quote": "Trained models of H-JEPA",
        "source_index": 3,
        "source_url": "https://huggingface.co/jepa-world-models/h-jepa",
        "observed_on": "2026-10-07",
        "snapshot_hash": "9d053b532805c979a09816ee4467e08a71a2016dbe5fdd042b6e81107efa49d3"
      }
    },
    "uniwam": {
      "route": {
        "value": "pixel",
        "quote": "visual prediction",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.02054v3",
        "observed_on": "2026-10-10",
        "snapshot_hash": "1c450a8f803c285f1ba721a38535622022db41ee1925ee3610e4c1f43a19e886"
      },
      "one_liner": {
        "value": "联合学习语义理解、未来视觉预测与动作生成。",
        "quote": "The hybrid model jointly supports semantic understanding, visual prediction, and action generation.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.02054v3",
        "observed_on": "2026-10-10",
        "snapshot_hash": "1c450a8f803c285f1ba721a38535622022db41ee1925ee3610e4c1f43a19e886"
      },
      "input": {
        "value": "当前观测、本体感知状态与任务指令。",
        "quote": "Given the current observation $o_{t}$, proprioceptive state $s_{t}$, and task instruction $I$",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.02054v3",
        "observed_on": "2026-10-10",
        "snapshot_hash": "1c450a8f803c285f1ba721a38535622022db41ee1925ee3610e4c1f43a19e886"
      },
      "output": {
        "value": "语言输出、未来观测与动作。",
        "quote": "the joint distribution of language outputs, future observations, and actions",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.02054v3",
        "observed_on": "2026-10-10",
        "snapshot_hash": "1c450a8f803c285f1ba721a38535622022db41ee1925ee3610e4c1f43a19e886"
      }
    },
    "deltaworld": {
      "route": {
        "value": "latent",
        "quote": "Delta-LTM predicts the action-conditioned latent increment",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.02691v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "a971d94074bfa505c38868f052a13fed71629bfed995ddfcc57d6ba717726335"
      },
      "one_liner": {
        "value": "预测动作引起的潜在状态增量，再更新状态并解码未来画面。",
        "quote": "Delta-LTM predicts the action-conditioned latent increment",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.02691v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "a971d94074bfa505c38868f052a13fed71629bfed995ddfcc57d6ba717726335"
      },
      "input": {
        "value": "初始视觉上下文与机器人动作序列。",
        "quote": "an initial visual context and a sequence of robot actions",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.02691v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "a971d94074bfa505c38868f052a13fed71629bfed995ddfcc57d6ba717726335"
      },
      "output": {
        "value": "解码后的未来 RGB 观测。",
        "quote": "Each predicted latent state is decoded by the frozen decoder to generate the corresponding RGB observation",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.02691v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "a971d94074bfa505c38868f052a13fed71629bfed995ddfcc57d6ba717726335"
      },
      "action_conditioned": {
        "value": true,
        "quote": "Delta-LTM predicts the action-conditioned latent increment",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.02691v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "a971d94074bfa505c38868f052a13fed71629bfed995ddfcc57d6ba717726335"
      }
    },
    "v2x-wam": {
      "route": {
        "value": "latent",
        "quote": "future occupancy and dynamic flow explicitly conditioned on the selected trajectory and action sequence",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2609.37098v2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "178ba44263e5b48f5eb909f9cac4eaaf08c172cc1aed6184f87027cedbbc91a4"
      },
      "one_liner": {
        "value": "以候选行驶轨迹与动作预测未来交通占据和动态流。",
        "quote": "future occupancy and dynamic flow explicitly conditioned on the selected trajectory and action sequence",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2609.37098v2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "178ba44263e5b48f5eb909f9cac4eaaf08c172cc1aed6184f87027cedbbc91a4"
      },
      "input": {
        "value": "车路协同场景表征、所选轨迹与动作序列。",
        "quote": "future occupancy and dynamic flow explicitly conditioned on the selected trajectory and action sequence",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2609.37098v2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "178ba44263e5b48f5eb909f9cac4eaaf08c172cc1aed6184f87027cedbbc91a4"
      },
      "output": {
        "value": "未来交通占据与动态流。",
        "quote": "future occupancy and dynamic flow explicitly conditioned on the selected trajectory and action sequence",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2609.37098v2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "178ba44263e5b48f5eb909f9cac4eaaf08c172cc1aed6184f87027cedbbc91a4"
      },
      "action_conditioned": {
        "value": true,
        "quote": "future occupancy and dynamic flow explicitly conditioned on the selected trajectory and action sequence",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2609.37098v2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "178ba44263e5b48f5eb909f9cac4eaaf08c172cc1aed6184f87027cedbbc91a4"
      }
    },
    "gwm-worlds-2": {
      "route": {
        "value": "interactive",
        "quote": "The model supports both first-person and third-person navigation",
        "source_index": 1,
        "source_url": "https://runway.com/research/introducing-gwm-worlds-2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "e03efa64bc17307caedd3a5d4075d1312ad8185e521dcd54f71b7fab77aba9a1"
      },
      "one_liner": {
        "value": "由文本动作与连续相机输入驱动的自回归视听世界模型。",
        "quote": "GWM Worlds 2 is an autoregressive diffusion video and audio model",
        "source_index": 1,
        "source_url": "https://runway.com/research/introducing-gwm-worlds-2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "e03efa64bc17307caedd3a5d4075d1312ad8185e521dcd54f71b7fab77aba9a1"
      },
      "input": {
        "value": "场景初始提示与首帧、当前相机和文本动作、历史生成帧。",
        "quote": "a global context (the genesis prompt and first frame), the current frame's inputs (camera input and the text actions that span the current frame) and the past generated frames",
        "source_index": 1,
        "source_url": "https://runway.com/research/introducing-gwm-worlds-2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "e03efa64bc17307caedd3a5d4075d1312ad8185e521dcd54f71b7fab77aba9a1"
      },
      "output": {
        "value": "逐步生成的视频与音频。",
        "quote": "GWM Worlds 2 is an autoregressive diffusion video and audio model",
        "source_index": 1,
        "source_url": "https://runway.com/research/introducing-gwm-worlds-2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "e03efa64bc17307caedd3a5d4075d1312ad8185e521dcd54f71b7fab77aba9a1"
      },
      "action_conditioned": {
        "value": true,
        "quote": "camera input and the text actions that span the current frame",
        "source_index": 1,
        "source_url": "https://runway.com/research/introducing-gwm-worlds-2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "e03efa64bc17307caedd3a5d4075d1312ad8185e521dcd54f71b7fab77aba9a1"
      },
      "limit": {
        "value": "快速相机旋转可能使细节、纹理和几何退化；长时程记忆仍不完善。",
        "quote": "Difficult camera inputs such as very quick rotations can cause details, textures and geometry to degrade. Long-term memory is also imperfect",
        "source_index": 1,
        "source_url": "https://runway.com/research/introducing-gwm-worlds-2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "e03efa64bc17307caedd3a5d4075d1312ad8185e521dcd54f71b7fab77aba9a1"
      }
    },
    "alayaworld-v1-1": {
      "route": {
        "value": "interactive",
        "quote": "viewpoint control is provided entirely through the re-rendered spatial condition",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2608.13492v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "7403e1f0ac521e4e0eeef918a7368c286956f9a28344484e468139f52c2ec660"
      },
      "one_liner": {
        "value": "以流式 3D 点缓存和重新渲染的空间条件生成视角可控视频。",
        "quote": "streaming 3D point-cache renderer",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2608.13492v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "7403e1f0ac521e4e0eeef918a7368c286956f9a28344484e468139f52c2ec660"
      },
      "input": {
        "value": "经重新渲染的空间记忆与运动感知视觉条件。",
        "quote": "viewpoint control is provided entirely through the re-rendered spatial condition",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2608.13492v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "7403e1f0ac521e4e0eeef918a7368c286956f9a28344484e468139f52c2ec660"
      },
      "output": {
        "value": "按视角条件生成的视频。",
        "quote": "temporal statistics consistent with those of the generated video",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2608.13492v1",
        "observed_on": "2026-10-07",
        "snapshot_hash": "7403e1f0ac521e4e0eeef918a7368c286956f9a28344484e468139f52c2ec660"
      },
      "public": {
        "value": "open_weights",
        "quote": "Pretrained weights v1.1",
        "source_index": 2,
        "source_url": "https://github.com/AlayaLab/AlayaWorld",
        "observed_on": "2026-10-07",
        "snapshot_hash": "956ae5479a4d934083ea0c0b79347c1877d5b72170de95c167dfc0e3b7217927"
      }
    },
    "alaya-evoke": {
      "route": {
        "value": "interactive",
        "quote": "external, camera-indexed world state bank",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2608.13546v2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "3e9ed7eac70fc08da52e91b32328cec0574651d2efc502f44a1e3409113ff651"
      },
      "one_liner": {
        "value": "借助外部世界状态库，持续生成受相机和文本条件控制的视频。",
        "quote": "external, camera-indexed world state bank",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2608.13546v2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "3e9ed7eac70fc08da52e91b32328cec0574651d2efc502f44a1e3409113ff651"
      },
      "input": {
        "value": "时间变化的文本、相机轨迹与投影历史。",
        "quote": "timed text schedule changes while the camera trajectory, warp history, and random seed remain fixed",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2608.13546v2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "3e9ed7eac70fc08da52e91b32328cec0574651d2efc502f44a1e3409113ff651"
      },
      "output": {
        "value": "交互生成的视频。",
        "quote": "interactive world model that generates 384 × 640 @ 24 fps video",
        "source_index": 3,
        "source_url": "https://huggingface.co/AlayaLab/Evoke",
        "observed_on": "2026-10-07",
        "snapshot_hash": "fc0bd99651baabb061cf5c953f980d01c4622e19bd0ac454eb6e3bf9d75ea2c1"
      },
      "public": {
        "value": "open_weights",
        "quote": "Model weights for EVOKE",
        "source_index": 3,
        "source_url": "https://huggingface.co/AlayaLab/Evoke",
        "observed_on": "2026-10-07",
        "snapshot_hash": "fc0bd99651baabb061cf5c953f980d01c4622e19bd0ac454eb6e3bf9d75ea2c1"
      },
      "limit": {
        "value": "固定预算的循环结构本身不能保证长时程内容一致。",
        "quote": "does not by itself ensure long-range content consistency",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2608.13546v2",
        "observed_on": "2026-10-07",
        "snapshot_hash": "3e9ed7eac70fc08da52e91b32328cec0574651d2efc502f44a1e3409113ff651"
      }
    },
    "inspatio-world-1-5": {
      "route": {
        "value": "interactive",
        "quote": "The target trajectory is still supplied by the user.",
        "source_index": 2,
        "source_url": "https://github.com/inspatio/inspatio-world-v1.5",
        "observed_on": "2026-10-07",
        "snapshot_hash": "26a047bd8f8b67038b54a17044e97b881df86fd9b0854e86e498b1d1bf56abb4"
      },
      "one_liner": {
        "value": "按用户提供的目标相机轨迹，从图像或视频生成视图视频。",
        "quote": "The target trajectory is still supplied by the user.",
        "source_index": 2,
        "source_url": "https://github.com/inspatio/inspatio-world-v1.5",
        "observed_on": "2026-10-07",
        "snapshot_hash": "26a047bd8f8b67038b54a17044e97b881df86fd9b0854e86e498b1d1bf56abb4"
      },
      "input": {
        "value": "单张图像、四张图像或视频，以及用户提供的目标相机轨迹。",
        "quote": "a single image, four images, or a video",
        "source_index": 2,
        "source_url": "https://github.com/inspatio/inspatio-world-v1.5",
        "observed_on": "2026-10-07",
        "snapshot_hash": "26a047bd8f8b67038b54a17044e97b881df86fd9b0854e86e498b1d1bf56abb4"
      },
      "output": {
        "value": "源画面、渲染画面、遮罩与预测视频文件。",
        "quote": "{source,render,mask,pred}.mp4",
        "source_index": 2,
        "source_url": "https://github.com/inspatio/inspatio-world-v1.5",
        "observed_on": "2026-10-07",
        "snapshot_hash": "26a047bd8f8b67038b54a17044e97b881df86fd9b0854e86e498b1d1bf56abb4"
      },
      "public": {
        "value": "open_weights",
        "quote": "Default v2v checkpoint — 1.3B",
        "source_index": 2,
        "source_url": "https://github.com/inspatio/inspatio-world-v1.5",
        "observed_on": "2026-10-07",
        "snapshot_hash": "26a047bd8f8b67038b54a17044e97b881df86fd9b0854e86e498b1d1bf56abb4"
      }
    },
    "solaris": {
      "route": {
        "value": "web",
        "quote": "Solaris generates the interface one frame at a time",
        "source_index": 1,
        "source_url": "https://runway.com/news/research/introducing-solaris",
        "observed_on": "2026-10-07",
        "snapshot_hash": "bf8346875782821c660f12ecc263c7e42b21bd6aea1e490b78980d395c085871"
      },
      "one_liner": {
        "value": "把用户交互作为条件，逐帧生成视觉界面。",
        "quote": "Solaris generates the interface one frame at a time",
        "source_index": 1,
        "source_url": "https://runway.com/news/research/introducing-solaris",
        "observed_on": "2026-10-07",
        "snapshot_hash": "bf8346875782821c660f12ecc263c7e42b21bd6aea1e490b78980d395c085871"
      },
      "input": {
        "value": "用户点击、拖拽等已发生的交互。",
        "quote": "The model observes clicks, drags and other interactions as it generates",
        "source_index": 1,
        "source_url": "https://runway.com/news/research/introducing-solaris",
        "observed_on": "2026-10-07",
        "snapshot_hash": "bf8346875782821c660f12ecc263c7e42b21bd6aea1e490b78980d395c085871"
      },
      "output": {
        "value": "逐帧生成的视觉界面。",
        "quote": "Solaris generates the interface one frame at a time",
        "source_index": 1,
        "source_url": "https://runway.com/news/research/introducing-solaris",
        "observed_on": "2026-10-07",
        "snapshot_hash": "bf8346875782821c660f12ecc263c7e42b21bd6aea1e490b78980d395c085871"
      },
      "action_conditioned": {
        "value": true,
        "quote": "The model observes clicks, drags and other interactions as it generates",
        "source_index": 1,
        "source_url": "https://runway.com/news/research/introducing-solaris",
        "observed_on": "2026-10-07",
        "snapshot_hash": "bf8346875782821c660f12ecc263c7e42b21bd6aea1e490b78980d395c085871"
      },
      "public": {
        "value": "application_required",
        "quote": "Fill out the form below to request early access.",
        "source_index": 1,
        "source_url": "https://runway.com/news/research/introducing-solaris",
        "observed_on": "2026-10-07",
        "snapshot_hash": "bf8346875782821c660f12ecc263c7e42b21bd6aea1e490b78980d395c085871"
      },
      "limit": {
        "value": "稳定、可读的文字仍是视频界面生成的困难。",
        "quote": "Stable, legible text remains one of the hardest problems in video generation",
        "source_index": 1,
        "source_url": "https://runway.com/news/research/introducing-solaris",
        "observed_on": "2026-10-10",
        "snapshot_hash": "1a17ccd75d02ea6e8f6c94b3037b7f1b4382882e97b782d7b137bd3cccdf9821"
      }
    },
    "odyssey-3": {
      "route": {
        "value": "interactive",
        "quote": "We build Odyssey-3 as a multi-step video diffusion transformer",
        "source_index": 1,
        "source_url": "https://odyssey.systems/meet-odyssey-3",
        "observed_on": "2026-10-10",
        "snapshot_hash": "8319c237e7d1e7e3591891b91f1c69a340eafc2fe71fb4694499d0d63b212461"
      },
      "one_liner": {
        "value": "以历史观察、提示和动作持续预测可交互的视觉环境。",
        "quote": "training it to continue from preceding observations and predict future states conditioned on action inputs.",
        "source_index": 1,
        "source_url": "https://odyssey.systems/meet-odyssey-3",
        "observed_on": "2026-10-10",
        "snapshot_hash": "8319c237e7d1e7e3591891b91f1c69a340eafc2fe71fb4694499d0d63b212461"
      },
      "input": {
        "value": "历史观察、文本提示和动作输入；预览支持第一/第三人称导航与独立相机移动。",
        "quote": "Today’s preview provides first-person and third-person navigation alongside independent camera movement",
        "source_index": 1,
        "source_url": "https://odyssey.systems/meet-odyssey-3",
        "observed_on": "2026-10-10",
        "snapshot_hash": "8319c237e7d1e7e3591891b91f1c69a340eafc2fe71fb4694499d0d63b212461"
      },
      "output": {
        "value": "随交互持续生成的未来环境视频。",
        "quote": "We build Odyssey-3 as a multi-step video diffusion transformer",
        "source_index": 1,
        "source_url": "https://odyssey.systems/meet-odyssey-3",
        "observed_on": "2026-10-10",
        "snapshot_hash": "8319c237e7d1e7e3591891b91f1c69a340eafc2fe71fb4694499d0d63b212461"
      },
      "action_conditioned": {
        "value": true,
        "quote": "predict future states conditioned on action inputs.",
        "source_index": 1,
        "source_url": "https://odyssey.systems/meet-odyssey-3",
        "observed_on": "2026-10-10",
        "snapshot_hash": "8319c237e7d1e7e3591891b91f1c69a340eafc2fe71fb4694499d0d63b212461"
      },
      "public": {
        "value": "limited_access",
        "quote": "You can try Odyssey-3 in our research preview",
        "source_index": 1,
        "source_url": "https://odyssey.systems/meet-odyssey-3",
        "observed_on": "2026-10-10",
        "snapshot_hash": "8319c237e7d1e7e3591891b91f1c69a340eafc2fe71fb4694499d0d63b212461"
      }
    },
    "dreamtrue": {
      "route": {
        "value": "pixel",
        "quote": "physically plausible video prediction.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.12468v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "55f54a12e5eb9028b4d239a33c7c52ceae91c2f2f4d06089350873d029459526"
      },
      "one_liner": {
        "value": "以机器人动作轨迹为条件，预测多视角未来视频。",
        "quote": "we render action trajectories into image-space conditions",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.12468v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "55f54a12e5eb9028b4d239a33c7c52ceae91c2f2f4d06089350873d029459526"
      },
      "input": {
        "value": "初始场景与机器人动作条件；轨迹经几何校准后渲染到图像空间。",
        "quote": "Predictions from the same initial scene under multiple action conditions.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.12468v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "55f54a12e5eb9028b4d239a33c7c52ceae91c2f2f4d06089350873d029459526"
      },
      "output": {
        "value": "多视角未来机器人操作视频。",
        "quote": "a multi-view, cross-embodiment robot world model for action-faithful and physically plausible video prediction.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.12468v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "55f54a12e5eb9028b4d239a33c7c52ceae91c2f2f4d06089350873d029459526"
      },
      "action_conditioned": {
        "value": true,
        "quote": "we render action trajectories into image-space conditions",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.12468v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "55f54a12e5eb9028b4d239a33c7c52ceae91c2f2f4d06089350873d029459526"
      },
      "limit": {
        "value": "动作表示、生成器和奖励模型均在图像空间工作；遮挡或视角不足时，可能生成或奖励看似合理但物理错误的交互。",
        "quote": "under occlusion or limited views, the framework may generate or reward visually plausible but physically incorrect interactions.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.12468v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "55f54a12e5eb9028b4d239a33c7c52ceae91c2f2f4d06089350873d029459526"
      }
    },
    "planwam": {
      "route": {
        "value": "latent",
        "quote": "A latent world model then predicts this planning-shaped future latent representation",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.11382v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "bb36833a204aa7b6b74f8c204a67d9b2ab8658bf74e5c84315f4d545f6e81709"
      },
      "one_liner": {
        "value": "从历史观察预测有助于驾驶规划的未来表征，再生成并选择轨迹。",
        "quote": "The predicted future latent representation serves as planning context and guides trajectory generation and selection.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.11382v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "bb36833a204aa7b6b74f8c204a67d9b2ab8658bf74e5c84315f4d545f6e81709"
      },
      "input": {
        "value": "历史多视角相机观察、自车状态与导航信息。",
        "quote": "denote the historical multi-view camera observations, $c_{t-n:t}$ the corresponding ego states and navigation information",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.11382v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "bb36833a204aa7b6b74f8c204a67d9b2ab8658bf74e5c84315f4d545f6e81709"
      },
      "output": {
        "value": "面向规划的未来潜在表征，以及由其引导生成和选择的自车轨迹。",
        "quote": "The predicted future latent representation serves as planning context and guides trajectory generation and selection.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.11382v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "bb36833a204aa7b6b74f8c204a67d9b2ab8658bf74e5c84315f4d545f6e81709"
      }
    },
    "long-wam": {
      "route": {
        "value": "pixel",
        "quote": "The video expert $\\theta$ predicts $K_{v}$ future latent steps",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.10528v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "640b0531e69c720101f5395a993b202759e95fc6fde1ebdbddfea05cf3f4619e"
      },
      "one_liner": {
        "value": "把长视觉历史与未来视频预测结合，用于生成机器人动作块。",
        "quote": "The resulting policy conditions actions on both observed history and anticipated futures",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.10528v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "640b0531e69c720101f5395a993b202759e95fc6fde1ebdbddfea05cf3f4619e"
      },
      "input": {
        "value": "观测视频历史、机器人状态与语言任务指令；具体格式随变体而定。",
        "quote": "observed video latents, $q_{t}$ the robot state, $c$ the language instruction",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.10528v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "640b0531e69c720101f5395a993b202759e95fc6fde1ebdbddfea05cf3f4619e"
      },
      "output": {
        "value": "未来视频潜变量与机器人动作块；IDM 先预测视频，再生成动作。",
        "quote": "it predicts future visual latents independently of action tokens, then denoises actions conditioned on those predictions",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.10528v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "640b0531e69c720101f5395a993b202759e95fc6fde1ebdbddfea05cf3f4619e"
      },
      "public": {
        "value": "open_weights",
        "quote": "Long-WAM policy weights for LIBERO.",
        "source_index": 3,
        "source_url": "https://huggingface.co/Efficient-Large-Model/Long-WAM-LIBERO-IDM",
        "observed_on": "2026-10-10",
        "snapshot_hash": "3b8a9350fe072fcbed05c881c584200834c7f3e22eeff92a63bf672e5b75aff5"
      }
    },
    "ppwm": {
      "route": {
        "value": "latent",
        "quote": "All outputs within a span are directly supervised by their corresponding future latent targets",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.08627v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "83d0762e5a60fe0c3daa0b615592fdf59473724d23a2992af97bd110f32f478a"
      },
      "one_liner": {
        "value": "在有限跨度内并行预测有因果关联的未来表征轨迹，用于动作规划。",
        "quote": "predicts a causal finite-horizon trajectory in parallel, allowing interaction among future representations",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.08627v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "83d0762e5a60fe0c3daa0b615592fdf59473724d23a2992af97bd110f32f478a"
      },
      "input": {
        "value": "近期潜在状态与已执行动作组成的上下文，以及待预测的动作序列。",
        "quote": "Given recent context and an action span",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.08627v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "83d0762e5a60fe0c3daa0b615592fdf59473724d23a2992af97bd110f32f478a"
      },
      "output": {
        "value": "有限跨度内的多步未来潜在表征轨迹。",
        "quote": "PPWM instead performs causal interaction among future hidden representations and decodes each future latent only after this interaction.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.08627v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "83d0762e5a60fe0c3daa0b615592fdf59473724d23a2992af97bd110f32f478a"
      },
      "action_conditioned": {
        "value": true,
        "quote": "Given recent context and an action span",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.08627v1",
        "observed_on": "2026-10-10",
        "snapshot_hash": "83d0762e5a60fe0c3daa0b615592fdf59473724d23a2992af97bd110f32f478a"
      }
    },
    "token-world": {
      "route": {
        "value": "latent",
        "quote": "model dynamics in a compact VLM visual-token space",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.00575v2",
        "observed_on": "2026-10-10",
        "snapshot_hash": "09e5d29c36ff5e95dd6f8855b00cfa0409b19541e323f03aa8eae49b647a1e9c"
      },
      "one_liner": {
        "value": "在策略使用的 VLM 视觉 token 空间中，按动作预测未来状态。",
        "quote": "an autoregressive action-conditioned world-model simulator that rolls out in compact policy-aligned VLM token space.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.00575v2",
        "observed_on": "2026-10-10",
        "snapshot_hash": "09e5d29c36ff5e95dd6f8855b00cfa0409b19541e323f03aa8eae49b647a1e9c"
      },
      "input": {
        "value": "压缩的 VLM 视觉 token、本体感知与机器人动作历史。",
        "quote": "Compact visual tokens, proprioception, and actions are projected into a shared model space",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.00575v2",
        "observed_on": "2026-10-10",
        "snapshot_hash": "09e5d29c36ff5e95dd6f8855b00cfa0409b19541e323f03aa8eae49b647a1e9c"
      },
      "output": {
        "value": "未来压缩 token 状态，可映射回策略使用的 VLM 视觉表征。",
        "quote": "Future compact tokens are generated with a sliding temporal window and recurrent state.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.00575v2",
        "observed_on": "2026-10-10",
        "snapshot_hash": "09e5d29c36ff5e95dd6f8855b00cfa0409b19541e323f03aa8eae49b647a1e9c"
      },
      "action_conditioned": {
        "value": true,
        "quote": "conditioned on the current history and action.",
        "source_index": 1,
        "source_url": "https://arxiv.org/html/2610.00575v2",
        "observed_on": "2026-10-10",
        "snapshot_hash": "09e5d29c36ff5e95dd6f8855b00cfa0409b19541e323f03aa8eae49b647a1e9c"
      }
    },
    "pixverse-r2": {
      "route": {
        "value": "interactive",
        "quote": "predicts the next synchronized audio-video segment from available history and live control.",
        "source_index": 1,
        "source_url": "https://pixverse.ai/en/blog/pixverse-r2-scaling-real-time-omni-world-models",
        "observed_on": "2026-10-10",
        "snapshot_hash": "de80059e75d7eaba183a5378385f140a3d209f10c19877ee3b3ac8e1a03546bb"
      },
      "one_liner": {
        "value": "以历史状态和实时多模态输入，持续生成同步音视频环境。",
        "quote": "predicts the next synchronized audio-video segment from available history and live control.",
        "source_index": 1,
        "source_url": "https://pixverse.ai/en/blog/pixverse-r2-scaling-real-time-omni-world-models",
        "observed_on": "2026-10-10",
        "snapshot_hash": "de80059e75d7eaba183a5378385f140a3d209f10c19877ee3b3ac8e1a03546bb"
      },
      "input": {
        "value": "历史环境状态，以及实时文本、参考材料、音频、WASD 等动作输入。",
        "quote": "Four types of input enter one running world: text prompts, multimodal references, audio, and actions such as WASD or continuous controls.",
        "source_index": 1,
        "source_url": "https://pixverse.ai/en/blog/pixverse-r2-scaling-real-time-omni-world-models",
        "observed_on": "2026-10-10",
        "snapshot_hash": "de80059e75d7eaba183a5378385f140a3d209f10c19877ee3b3ac8e1a03546bb"
      },
      "output": {
        "value": "连续更新的同步视频与音频片段。",
        "quote": "predicts the next synchronized audio-video segment from available history and live control.",
        "source_index": 1,
        "source_url": "https://pixverse.ai/en/blog/pixverse-r2-scaling-real-time-omni-world-models",
        "observed_on": "2026-10-10",
        "snapshot_hash": "de80059e75d7eaba183a5378385f140a3d209f10c19877ee3b3ac8e1a03546bb"
      },
      "action_conditioned": {
        "value": true,
        "quote": "They can arrive while the model is generating, alter the next world state, and drive synchronized video and audio.",
        "source_index": 1,
        "source_url": "https://pixverse.ai/en/blog/pixverse-r2-scaling-real-time-omni-world-models",
        "observed_on": "2026-10-10",
        "snapshot_hash": "de80059e75d7eaba183a5378385f140a3d209f10c19877ee3b3ac8e1a03546bb"
      }
    }
  }
}