Transcript · English → Chinese
Training Robots in Virtual Environments
机器人目前正在“母体” 或类似的环境中接受训练。
让我来解释一下。
英伟达开发了名为 IsaacSim 的平台， 它是一个超逼真的虚拟环境
机器人可以在其中练习行走、捡起物 品，
并在无需任何实体机器人的情况下 进行空间导航。
他们还开发了名为 NVIDIA Cosmos 的系统，
这是一个可以根据需求 生成逼真环境的世界基础模型。
因此，从厨房到仓库再到工厂车间，
Cosmos 都能构建出符合真实世界 光照和物理规律的场景。
在这些生成的虚拟世界中， 有一群虚拟机器人正在接受训练。
而且这不仅仅是每次训练一个。
而是有成千上万个相同机器人的副 本 在并行运行，
而且每一个所处的环境都略有不同。
现在，由于模拟运行速度快于实时，
机器人可以在短短一天内 体验相当于十年的训练。
一旦机器人在模拟中学会了某项技 能，
该技能就会被加载到 一个名为Groot的基础模型中，
这是英伟达为人形机器人开发的大 脑。
然后，这个大脑可以被下载到 这些实体机器人中的任何一个
并部署到现实世界中。
因此，当机器人第一次走出仓库时，
它就已经拥有了 数千个生命周期的经验。
End of transcript · 239 words · machine translated
Original transcript (English)
Robots are currently being trained in the Matrix, or something just like it. Let me explain. NVIDIA built a platform called IsaacSim, which is a hyper-realistic virtual environment where robots can practice walking, picking things up, and navigating spaces without any physical robots involved.
They also built something called NVIDIA Cosmos, which is a world foundation model that can generate realistic environments on demand. So anything from kitchens to warehouses and factory floors, Cosmos can build any of them with realistic lighting and physics that match the real world. Inside those generated worlds, there is a bunch of virtual robots training.
And it's not just one at a time. There's thousands of copies of the same robot running in parallel, and each one is experiencing slightly different conditions. Now, because the simulation runs faster than real time, a robot can experience the equivalent of 10 years of training in a single day.
Once the robots learn a skill in the simulation, that skill gets loaded into a foundation model called Groot, which is NVIDIA's brain for humanoid robots. And then that brain can be downloaded into any of these physical robots and deployed in the real world.
And so when a robot walks out of a warehouse for the first time, it already has thousands of lifetimes of experience.
