NVIDIA Research Unveils Advanced Grasping, Smart Autonomous Driving, and Large-Scale Agent Training Technologies

At CVPR, NVIDIA Research presented three physical AI research achievements that can be broadly applied to various applications through large-scale training.

GraspGen-X is the first foundation model for zero-shot grasping. Trained on 2 billion simulated grasp data points, it understands geometric structures and generates stable grasp poses for novel gripper shapes or previously unseen objects without any additional retraining.

LCDrive innovates the inference speed of autonomous driving systems. While existing text-based chain-of-thought approaches are constrained by token generation time on in-vehicle embedded hardware, LCDrive replaces this with compressed latent representation, enabling much faster decision-making.

NitroGen is a general-purpose gameplay AI foundation model that leverages the NVIDIA Isaac GR00T architecture. It supports agent learning through tens of thousands of hours of virtual environment interactions.

NVIDIA also unveiled new physical AI agent skills to help researchers and developers more quickly develop autonomous driving, robotics, and vision AI systems.