Details
- NVIDIA Research presented three CVPR papers showing how large-scale training across grippers, driving scenarios and virtual worlds improves AI generalization for physical and embodied agents.
- The work centers on GraspGen-X, LCDrive and NitroGen, along with newly unveiled physical AI agent skills to accelerate autonomous vehicle, robotics and vision AI development.
- GraspGen-X is described as the first foundation model for zero-shot grasping, trained on 2 billion simulated grasps across thousands of object shapes and synthetic grippers to generate grasp poses for virtually any gripper, and can be paired with the curoboV2 CUDA motion-planning library.
- LCDrive replaces text-based chain-of-thought with compact spatial latent representations, halving token usage while maintaining trajectory quality so autonomous vehicles can reason faster on embedded hardware, building on NVIDIA’s Alpamayo autonomous driving platform.
- NitroGen adapts the Isaac GR00T robot foundation model architecture to train agents over 40,000 hours across more than 1,000 video games, improving low-data performance by up to 52% versus prior methods and is released open source on GitHub and Hugging Face.
Impact
These releases push physical AI toward foundation-style models that generalize across embodiments, from robot grippers to game-trained virtual agents to automotive stacks. By focusing on scalable simulation data and efficient on-device reasoning, NVIDIA strengthens its position at the intersection of AI models and specialized hardware, likely steering robotics, AV, and embodied-agent R&D toward large, multi-domain training regimes over the next 12–24 months.
