In the announcement of Claude Fable 5.2, Anthropic specifically boasted about how the model controls robots. Fable 5.2 performs significantly better with physical robots, even in tasks it wasn't specifically trained for. On the RoboDojo benchmark with a simulated two-armed robot, it outperformed other models and most AI systems designed specifically for robot control. In the video, Claude assembles a puzzle cube, unscrews a bottle cap, and inserts a gear, checking the engagement.
I recently wrote about GR00T 2 from NVIDIA - a specialized model trained on hundreds of thousands of hours of videos with humans and robots. This is one path: building a separate "brain" for robots. But here we see that a universal language model, which wasn't taught to unscrew caps, already surpasses most of these specialized systems.
I think it might well turn out that the GPT moment in robotics, which I am waiting for, will happen thanks to general models: they will become so intelligent that the physical world will be just another task for them, without a separate perfect dataset for robots.
https://x.com/claudeai/status/2119867132452221045
#robotics@rvnikita_blog #anthropic@rvnikita_blog #claude@rvnikita_blog #ai@rvnikita_blog