NVIDIA Launches Open Source Physical AI Tools, Pegatron Cuts Training 67%, Unveils Jetson Thor Humanoid Robot
NVDA•NVIDIA launched open source physical AI agent tools for Omniverse, Cosmos, Alpamayo and Metropolis used by Pegatron and Foxconn—with Pegatron cutting training time by 67%. It also unveiled the Isaac GR00T humanoid reference robot on Jetson Thor delivering 2,070 FP4 teraflops and 128GB memory to accelerate academic research.
1. Open Source Physical AI Agent Tools
NVIDIA released a comprehensive open source toolkit within its Agent Toolkit that enables AI agents to orchestrate physical AI workflows across Omniverse, Cosmos, Alpamayo and Metropolis. The suite includes agent-callable instructions for data generation, simulation, training and deployment, and has been adopted by partners such as Pegatron, which cut model training time by 67%, along with Foxconn and TSMC.
2. Isaac GR00T Humanoid Robot Reference Design
NVIDIA introduced its Isaac GR00T reference humanoid robot, integrating a Unitree H2 Plus chassis and Sharpa Wave five-fingered hands with NVIDIA Jetson Thor compute delivering 2,070 FP4 teraflops and 128GB unified memory. The open design supports whole-body control, multi-view sensing and extended battery life, aiming to accelerate academic research at institutions like Stanford and ETH Zurich.


