Nvidia has unveiled GR00T 2, an open AI model for teaching robots precise hand movements using human video. Nvidia is making 100,000 hours of first-person video used in training available for anyone to use.
Nvidia describes GR00T 2 as a world action model: it predicts how an action will change what a robot sees before it moves. Trained on robot data and human video, it can transfer skills between robots with different bodies, according to Nvidia.
In Nvidia's tests, the model achieved 71% average success after seeing a task demonstrated once, rising to 95% after five demonstrations. However, the company's paper says the tasks were selected from ones the model could already learn well.
Foxconn is using GR00T 2 to help build AI servers, while Siemens is testing factory applications and Nexuni is trying laundry tasks. None has announced production use. For robotics builders, the model and footage offer development resources, but the reported results do not establish reliability on unfamiliar commercial tasks.
The open model and video release give robotics builders new resources for developing systems that handle physical objects.
Learning skills from human footage could reduce dependence on demonstrations collected separately for each robot body.
Commercial reliability remains unproven because the tests used selected tasks and the named partners have not announced production use.
Written automatically by the Made in 10 Newsroom from the source above.
Meta Platforms is expanding AI age estimation worldwide to place more Instagram users in protected Teen Accounts. The checks can apply even when users have supplied birth dates showing they are adults.
Google is introducing TPU 8tx and 8i Max, specialised variants of its AI chips. The variants target confidential computing and AI applications that need to process more information in a single request.
Uber engineers say they changed how code is merged after AI-written changes overwhelmed its existing process. They report that allowing more parallel builds brought peak throughput to almost 250 and sometimes 300 merges an hour, but introduced new testing risks.