Nvidia launched GR00T 2 at CoRL on November 10, 2026, a robot foundation model designed to control arms, hands, and humanoid robots. The model pretrains on human video rather than relying primarily on expensive teleoperated robot recordings. Sources give different figures for training data: one states 100,000 hours of human egocentric data, while another states 402,730 hours of human head-camera video plus 10,184 hours of handheld-gripper demos and 23,500 hours of teleoperated robot data.
GR00T 2 completed 40 of 50 evaluated tasks at least once without demonstrations, achieves 71% average success with one demonstration per task, and 95% with five demonstrations. The model ranks among the top three across all vision language action and world action model benchmark categories. Running in Turbo mode on Jetson Thor hardware, it achieves 124 ms per prediction while planning 1.6-second chunks of whole-body motion.
Nvidia describes GR00T 2 as open, but sources diverge on model weights: one reports weights and training dataset will be released in December 2026, while another reports no confirmed release date or license for model weights. Code availability is listed for December. The human video training dataset is already available.
Foxconn is applying GR00T 2 to server-manufacturing automation, Nexuni is testing it for robotic laundry handling, MicroAGI is using GR00T 2 post-trained via reinforcement learning for Ethernet cable swapping with a YAM robotic arm, and Siemens is evaluating it for industrial uses. LG Electronics and Nvidia demonstrated Isaac GR00T-powered robotic manipulation in a live LG manufacturing environment using LG's CLOiD robot, implementing a physical AI data flywheel connecting real-world data collection, Cosmos-based data augmentation, GR00T model training, Jetson AGX Thor onboard compute, and simulation-based evaluation.
Isaac Lab 3.0 adds Newton physics engine integration enabling simulation of deformable objects such as cables and cloth, and delivered up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX in a vision-based dexterous manipulation benchmark. Skild AI announced STEEL, a rigid-body physics solver built on Newton. FieldAI, Dyna Robotics, Flexion, and Walden Robotics are among partners using Isaac Lab and Newton.
Three new benchmarks were introduced at CoRL: Isaac Lab-Arena 0.4 enables parallel policy testing across thousands of condition variations; DexBench, developed by RLWRLD and Nvidia, covers 15 tasks across retail, logistics, manufacturing, and hospitality; and VABAR, from UC Berkeley and Nvidia, evaluates robots across six industrial tasks with standardized difficulty tiers. UC Berkeley's Graph-as-Policy framework chains individual robot skills into programs that AI agents test and refine in simulation before physical deployment. Ambi Robotics, Bosch Research, and Siemens are exploring GaP for industrial automation workflows.
AlphaNIST achieved more than 98% training success on all 20 NIST Assembly Task Board tasks with a single policy, compared with 81% for separate per-task policies, using approximately 39x less GPU time. Two Nvidia papers, GRAIL and RoboTTT, are among six Best Paper Award finalists at CoRL; RoboTTT scales visuomotor context three orders of magnitude beyond prior policies without added inference latency.