跳到主要内容Skip to main content

← 全部作品← All pieces / 模型演示Model demos / 203

模型演示Model demos203 / 210

Gemini 4 教人形机器人投沙包Teaching a humanoid robot to play cornhole with Gemini 4 Argon

模型:Gemini 4 Argon。原始发布页、源码、prompt 与许可都在下方。Model: Gemini 4 Argon. Original publisher page, source, prompt and licence are below.

作品在原作者的页面运行。如果画面空白,通常是对方禁止嵌入,点“全屏打开原页”。The piece runs on the original author’s page. If it appears blank, embedding is likely blocked; use “Open original”.

这件作品About this piece

演示什么:Google DeepMind 的机器人团队让 Gemini 4 Argon 看一段人投沙包的视频,模型在 MuJoCo 里搭出虚拟仿真,算出物理与全身控制参数,再交给 Apptronik 的 Apollo 2 人形机器人执行;官方称它在真实世界连续投进 7 球。 官方怎么说:这条 Video-to-Sim-to-Real 流水线把「几周的手工建模」压到「几小时」,目标是让普通视频变成物理 AI 的训练场。 边界:这是内部测试的单个任务,不是跨任务成功率;视频没有给出尝试次数、失败分布、 sim-to-real 偏差或安全边界,不能外推成「人形机器人已通用」。What it shows: Google DeepMind robotics researchers had Gemini 4 Argon watch a video of a person playing cornhole; the model built a MuJoCo simulation, calculated the physics and whole-body control, and handed them to Apptronik’s Apollo 2 humanoid, which the official video says sank seven consecutive shots in the real world. What the official video says: this Video-to-Sim-to-Real pipeline compresses weeks of manual modelling into hours, turning ordinary videos into a training ground for physical AI. Boundary: this is a single internal test task, not a cross-task success rate; the video gives no attempt count, failure distribution, sim-to-real gap or safety envelope, so it does not generalise to “humanoids are solved”.

来源与 promptSource & prompt

Google 官方发布内容;本站只链回官方视频或原文,不复制素材。引用与转载以 Google 的条款为准。Official Google release; this site only links to the official video or article and does not copy the media. Quotation and redistribution follow Google’s terms.