NVIDIA Research Unveils GraspGen-X, LCDrive, and NitroGen for Smarter Autonomous Driving and Robot Grasping
NVIDIA presented three AI models at CVPR that enable robots to grasp novel objects, autonomous vehicles to reason faster on embedded hardware, and virtual agents to train across thousands of game environments.
- 2 billion
- 40,000 hours
- up to 52%
What Happened
At CVPR 2026, NVIDIA Research presented three papers—GraspGen-X, LCDrive, and NitroGen—that tackle key challenges in physical AI: zero-shot robotic grasping, efficient reasoning for autonomous vehicles, and scalable training of embodied agents. These models share a common theme of training at scale to generalize across diverse applications.
2 billiongrasps
Generated across thousands of object shapes and gripper configurations, enabling the first foundation model for zero-shot grasping.
LCDrive replaces text-based chain-of-thought reasoning with compact latent representations, allowing autonomous vehicles to think faster on embedded hardware. It achieves comparable trajectory quality to text-based reasoning using roughly half the tokens.
52%improvement
Over previous state-of-the-art methods for embodied agent training.
- GraspGen-X: first foundation model for zero-shot grasping, works with any gripper.
- LCDrive: latent reasoning reduces token count by half for autonomous driving.
- NitroGen: trained on 1,000+ games and 40,000 hours of interaction, open source.
Why this matters
These advances bring physical AI closer to real-world deployment by making robotic grasping more flexible, autonomous driving more efficient, and AI agent training more scalable.
Terms in This Story
- Foundation model
- A large AI model trained on broad data that can be adapted to a wide range of tasks.
- Zero-shot grasping
- The ability of a robot to grasp objects it has never seen before without additional training.
- Latent representation
- A compressed encoding of information that captures essential features, used to reduce computational load.
- Chain-of-thought reasoning
- A prompting technique that breaks down a problem into intermediate steps to improve AI decision-making.
Summarised from the linked release; details can be imperfect — always verify against the original source.