NVIDIA's Innovations in Robotics Simulation at CoRL

During the Conference on Robot Learning (CoRL) in Austin, Texas, NVIDIA introduces new open technologies for robotics, focused on refining robot manipulation. While traditional applications in simulation were primarily concerned with locomotion, the emphasis is now on performing complex tasks, such as aligning mechanical parts and assembling small components.

NVIDIA presents in this context the GR00T 2, an advanced world action model that enables robots to respond accurately and perform multistep tasks. This model has made a significant leap forward in terms of skill transfer and generalization, which will add robustness to robotics in industrial environments, where precision and adaptability are crucial.

In collaboration with various partners, including Bosch, Siemens, and UC Berkeley, NVIDIA is working to implement these innovations within the broader robotics community. In addition to GR00T 2, new benchmarks and evaluation frameworks are being introduced, such as DexBench and VABAR, designed to measure the skill and adaptability of robots in industrial tasks.

Another important development is the agentic robotics approach, where individual robot skills are combined into coordinating programs. This makes it easier for robots to adapt to changing circumstances and increases efficiency in the development process. With these technologies, researchers can optimize the training of robots and effectively evaluate their performance.

Read the full article from Nvidia.