由Google Cloud赞助
选择您的第一个基因AI 使用案例
要从基因AI开始，首先关注能够改善人类信息体验的领域。
更新版指出机器人和AI测试中世界模型的新用途，但企业必须权衡其潜力与准确性差距。
4 读取
Romolo Tavani通过 Getty 图像
企业总有一天可以在生成的世界中排练机器人和AI agents，然后在工厂地板上冒险，但Google DeepMind的"Genie 4"中的准确性差距限制了公司今天能够采取这个方法的多远。
发布于10月27日作为有限的研究预览，更新后的世界模型实时生成交互设置，让用户随着场景的展开而改变。
Genie 4以它的前身"Genie 3"为基础，增加了本土的音频和对话，更现实的物理和更自然的人类运动。 用户可以通过改变天气，添加对象或字符并随着场景的展开来调整相机角度来使用自然语言的提示来修改环境。
研究人员可以利用这些生成的环境来探索包括山洪暴发和野火等自然灾害是如何展开的。 它们还可以创造户外情景，以测试自主车辆、无人驾驶飞机和其他机器人系统如何应对不断变化的天气和环境条件。
一个世界模型通过帧生成一个交互式世界框架，reacts到用户或AI agents做什么。 在与Gemini机器人相关的研究设置中，Google将Genie 4与Gemini模型配对，这些模型将任务分解为步骤，监测模拟场景并随着条件变化而调整计划。
称是AI研究与策略公司的创始人兼首席执行官
"它可以给企业一种创造和测试不同情景的方法，而不必在现实世界中重新创造这些情景"，他称。
Google最清晰的企业应用是机器人。 在一次演示中，Google使用Genie 4来生成一个模拟Apptronik的阿波罗人形机器人扮演检查器的设定。 演示表明，公司最终如何利用类似情况在将机器人部署到工厂地上或仓库之前通过各种情况运行机器人。 开发者可以将机器人软件暴露在不同的布局，障碍或照明条件下而无需实际构建出每一种情景。
"企业团队最终可以利用生成的环境来原型体验，创建培训情景并测试AI agents在现实世界中成本高或难以再生的情况","战略地平线研究的创始人兼CEO称 Manish Jain，一家AI战略和咨询公司。
塞兰德还看到了快速原型化的潜力，包括建造交互式相机和测试相机角度，照明和布局。
"世界模型可以帮助解决代理人AI的局限性，它主要通过文本和API运行：在模拟环境中测试一个动作的物理后果的能力"，他称。
该技术仍为实验技术，而Google承认其局限性。 物体和视觉细节可以漂移，特别是在复杂的即时变化后或当用户导航环境时。 该公司还列举了复杂的物理互动、空间准确性、音频准确性、可辨认文本和长期一致性方面的挑战。
这些差距很重要，因为机器人培训必须足够一致，使开发者能够比较结果并了解系统成功或失败的原因。 如果模拟条件不能准确反映物理世界，开发者可能难以确定机器人的性能是否会转移到现实世界的设定。
这被称为机器人学中的现实差距。 2026年对现实差距的回顾考察了模拟和物理设置之间的差异如何能阻止机器人在模拟之外按预期进行表演。
Selland警告说Genie 4仍然是早期的研究预览。 "会议可以漂移，并且它的物理仍然大致，使它不适合今天的工程环境需要精确。"
Jain同样认为，关键的问题是，结果的转移是否超出模拟范围。 "在依靠所生成的环境来训练机器人或验证一个物理过程之前，企业需要证明它的条件和结果充分符合真实的任务"，他称。
企业在依靠这些模型完成艰巨任务之前，应先在沙箱中广泛运行世界模型，并补充说，如果不在内部充分评价这些模型，它们就不大可能信任这些模型，Selland 称。
Google计划通过它的消费原型"Genie项目"向Google AI Ultra用户提供Genie 4，同时向Gemini API和Google AI Studio的早期接入合作伙伴提供。
有限的推出仍然限制了企业开发者的访问。 Google没有公布定价或具体发布日期。
其他公司也在发展世界模式。 Runway 发布它的可玩世界模型"GWM Worlds 2"配有音频，于9月出道；而由AI先驱费-费·李所创立的世界实验室也在现场工作。 向物理AI和机器人的更广阔的推进也促使人们对模拟培训和测试感兴趣。
几个卖家竞相上市买家的疑问是哪家声称现在行动。
Jain 称企业可以首先使用世界模型来创建虚拟场景来进行原型化和培训AI系统。 但是，利用这些环境来训练机器人或验证机器人在物理世界中将如何表现，是一个要求更高的应用，需要更高的模拟精度。
"商业界应该区分使用世界模型来实验情景和依靠它们来预测机器人或物理过程在现实世界中的行为方式". He 称。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
The updated version points to new uses for world models in robotics and AI testing, but enterprises must weigh their potential against accuracy gaps.
4 Min Read
Romolo Tavani via Getty Images
Enterprises could someday rehearse robots and AI agents in generated worlds before risking them on a factory floor, but accuracy gaps in Google DeepMind's Genie 4 limit how far companies can take that approach today.
Released on Oct. 27 as a limited research preview, the updated world model generates interactive settings in real time, letting users change them as a scene unfolds.
Genie 4 builds on its predecessor, Genie 3, adding native audio and dialogue, more realistic physics and more natural human movement. Users can modify environments using natural-language prompts by changing the weather, adding objects or characters and adjusting camera angles as scenes unfold.
Researchers can use these generated settings to explore how natural disasters, including flash floods and wildfires, unfold. They could also create outdoor scenarios to test how autonomous vehicles, drones and other robotic systems respond to changing weather and environmental conditions.
A world model generates an interactive world frame by frame and reacts to what users or AI agents do. In a research setup tied to Gemini Robotics, Google paired Genie 4 with Gemini models that break tasks into steps, monitor the simulated scene and adjust plans as conditions change.
Chris Selland, founder and CEO of Differential Factor, an AI research and strategy company, said the important difference is that Genie 4 can create an interactive space that responds to what a person or AI system does, rather than simply generating a video from a prompt.
"It could give businesses a way to create and test different scenarios without having to recreate them in the real world," he said.
Google's clearest enterprise application is robotics. In a demonstration, Google used Genie 4 to generate a setting in which a simulated Apptronik’s Apollo humanoid robot played checkers. The demonstration illustrates how companies could eventually use similar situations to run robots through varied scenarios before deploying them on a factory floor or in a warehouse. Developers could expose robot software to different layouts, obstacles or lighting conditions without having to physically build each scenario.
"Enterprise teams could eventually use generated environments to prototype experiences, create training scenarios and test AI agents in situations that are costly or difficult to reproduce in the real world," said Manish Jain, founder and CEO of Strategic Horizon Research, an AI strategy and advisory firm.
Selland also sees potential in rapid prototyping, including building interactive sets and testing camera angles, lighting and layouts.
"World models could help address a limitation of agentic AI, which has largely operated through text and APIs: the ability to test the physical consequences of an action in a simulated environment," he said.
The technology is still experimental, and Google acknowledges limitations. Objects and visual details can drift, especially after complex prompt changes or as users navigate an environment. The company also cites challenges with complex physical interactions, spatial accuracy, audio accuracy, legible text and long-term consistency.
Those gaps matter because robot training must be sufficiently consistent for developers to compare results and understand why a system succeeds or fails. If simulated conditions don't accurately reflect the physical world, developers could have difficulty determining whether a robot's performance will transfer to a real-world setting.
This is known as the reality gap in robotics. A 2026 review of the reality gap examines how differences between simulated and physical settings can prevent robots from performing as expected outside a simulation.
Selland cautioned that Genie 4 is still an early research preview. "Sessions can drift, and its physics remain approximate, making it unsuitable today for engineering environments that require precision."
Jain similarly argued that the critical question is whether results transfer beyond the simulation. "Before relying on a generated environment to train robots or validate a physical process, enterprises would need evidence that its conditions and outcomes sufficiently match the real task," he said.
Enterprises should run world models extensively in sandboxes before relying on them for demanding tasks, adding that they are unlikely to trust the models without fully evaluating them internally, Selland said.
Google plans to make Genie 4 available to Google AI Ultra subscribers through Project Genie, its consumer prototype, and to early-access partners in the Gemini API and Google AI Studio.
The limited rollout still restricts access for enterprise developers. Google has not published pricing or specific release dates.
Other companies are also developing world models. Runway released its playable world model, GWM Worlds 2, with audio, in September, while World Labs, founded by AI pioneer Fei-Fei Li, is also working in the field. The broader push toward physical AI and robotics is also driving interest in simulation-based training and testing.
With several vendors racing to market, the question for buyers is which claims to act on now.
Jain said enterprises could begin by using world models to create virtual scenes for prototyping and training AI systems. But using those environments to train robots or verify how they will perform in the physical world is a more demanding application that requires much greater simulation accuracy.
"Businesses should distinguish between using world models to experiment with scenarios and relying on them to predict how robots or physical processes will behave in the real world," he said.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.