NVIDIA open-sources GR00T 2 and 100K hours of robot data
NVIDIA put the weights, the simulator, the hardware designs and 100,000 hours of training data on the table today. Google DeepMind, meanwhile, says Apollo 2 humanoids are already kitting car parts at Mercedes-Benz with 96% accuracy.
Shown this week
NVIDIA opens GR00T 2: weights, sim engine and 100K hours of data
The GR00T 2 release landed in full, and the giveaway is bigger than the model: Jim Fan says NVIDIA is publishing weights, agentic framework, simulation engine, annotation engine, hardware designs and 100,000 hours of training data.
It runs entirely on your robot’s onboard compute, fully untethered, and you own every weight of it.
Gemini Robotics 2.5 kits car parts at Mercedes with 96% accuracy
Google DeepMind released Gemini Robotics 2.5 and put it on Apptronik's Apollo 2 humanoids inside Mercedes-Benz plants, where Carolina Parada says the robots picked and organized car parts for assembly with 96% accuracy.
As a researcher it is quite surreal to see robotics go from not working at all, to inventing VLAs, to now real-world usage across the world
Claude's Fable 5.2 beats robot-specific models on a two-arm benchmark
Anthropic says Fable 5.2 outperformed other models and most AI systems built to control robots on RoboDojo, a benchmark with a simulated two-arm robot, including on tasks it was never explicitly trained to do. Jiafei Duan is running it head to head against other policies live at CoRL2026, at the BitRobot booth for RoboColosseum.