NVIDIA Unveils GraspGen-X Foundation Model for Zero-Shot Robotic Grasping at CVPR 2026

Summary

NVIDIA Research is presenting three groundbreaking papers at CVPR 2026 focused on physical AI, with a major emphasis on GraspGen-X, the first foundation model for zero-shot robotic grasping. The model was trained on billions of simulated grasps, enabling robots to grasp novel objects without task-specific fine-tuning. These research contributions represent significant advances in scaling training techniques across diverse physical AI applications, addressing one of the most critical challenges in robotics: generalizable object manipulation across varied scenarios and environments.

The release of GraspGen-X demonstrates NVIDIA's strategic focus on bridging the gap between simulation and real-world robotics through foundation models. By leveraging vast synthetic training data, the model achieves unprecedented generalization capabilities, allowing robots to handle real-world grasping tasks they've never explicitly seen during training. This work is part of a broader set of three papers that collectively advance the state-of-the-art in training large-scale physical AI systems.

• Foundation models for robotics enable real-world deployment of AI systems trained entirely in simulation

Editorial Opinion

GraspGen-X represents a watershed moment in robotics—the successful application of foundation model scaling to physical manipulation. This approach mirrors the breakthroughs seen in language and vision models, suggesting that robotics has entered a new era where large-scale pretraining on synthetic data can unlock remarkable real-world capabilities. The work validates NVIDIA's bet on simulation and synthetic data as a path to practical AI robotics, with significant implications for manufacturing, logistics, and warehouse automation.