\n
![](\"https:\/\/blogs.nvidia.com\/wp-content\/uploads\/2026\/06\/CVPR_blog_still-scaled.png\")
<\/p>\n
\n<\/p>\n
What makes a robot gripper useful isn\u2019t that it can pick up one object \u2014 it\u2019s that it can pick up the next one, and the one after that, with a tool it\u2019s never held before.\u00a0<\/span><\/p>\n What makes an autonomous vehicle system safe isn\u2019t just that it can reason through a situation \u2014 it\u2019s that it can do so quickly enough on the hardware actually installed in the car.\u00a0<\/span><\/p>\n What makes a virtual agent capable is exposure to as many different environments as possible before it faces the real world.\u00a0<\/span><\/p>\n At this year\u2019s Computer Vision and Pattern Recognition (CVPR) conference, NVIDIA Research is presenting three papers that address each of these challenges \u2014 and share a common theme: training at scale creates systems that generalize across diverse applications.<\/span><\/p>\n The three papers cover different challenges in physical AI research:\u00a0<\/span><\/p>\n NVIDIA also unveiled at CVPR <\/span>new physical AI agent skills<\/span><\/a> that help researchers and developers speed the development of autonomous vehicles, robots and vision AI systems.<\/span><\/p>\n NitroGen and another NVIDIA-authored paper, PixelDIT<\/a>, were named best paper finalists at the conference \u2014 an accolade given to just 15 of over 4,000 accepted papers at CVPR.<\/p>\n Most AI systems for robotic grasping are specialists.<\/span><\/p>\n A <\/span>vision-language-action<\/span><\/a> policy trained for a two-finger gripper only learns to grasp with those two fingers. Similarly, a policy for dextrous grasping will only work for the bespoke multi-fingered gripper it\u2019s trained on. For every new embodiment, the process typically needs to be repeated \u2014 requiring new training data, fine-tuning and validation. This constraint means most robotics companies pick a gripper, train for it and stick with it.<\/span><\/p>\n GraspGen-X<\/b><\/a> is the first foundation model for grasping built to eliminate this bottleneck.\u00a0<\/span><\/p>\n Like a large language model that can apply its understanding of language to a new task without retraining, GraspGen-X applies its understanding of geometry and contact to any robotic gripper it encounters. Given the geometry of a new gripper and an unknown object it\u2019s never seen before, the model generates reliable grasp pose proposals to enable the robot to grasp the object.<\/span><\/p>\n To get there, the researchers needed a dataset that\u2019s impossible to collect in the real world at scale. They generated 2 billion simulated grasps across thousands of object shapes and synthetic gripper configurations, spanning the diversity of form factors a deployed robot might encounter.\u00a0<\/span><\/p>\n For robot developers, this foundation model eliminates the need for per-gripper training cycles and can be applied out of the box for several commonly used grippers. GraspGenX can be used in conjunction with <\/span>curoboV2<\/span><\/a>, a new CUDA-accelerated motion planning library, to achieve these grasp poses in unknown environments.\u00a0<\/span><\/p>\n Building on the GraspGen research foundation, another paper, <\/span>Grasp-MPC \u2014 presented at ICRA 2026<\/span><\/a> \u2014 advances the next step in the pipeline: moving from grasp generation to closed-loop grasp execution.<\/span><\/p>\n In recent years, researchers have found that letting an AI reason \u2014 generating intermediate thinking steps before committing to an answer \u2014 reliably improves its decision-making.\u00a0<\/span><\/p>\n For autonomous vehicles, the challenge is doing that reasoning on the hardware inside an actual vehicle. Text-based chain-of-thought reasoning generates words, and every word is a token that takes time to produce. On the processor running inside a car, token count is a real constraint on how fast the system can respond.<\/span><\/p>\n LCDrive<\/b> tackles this problem by replacing words with compressed latent representations.\u00a0<\/span><\/p>\n Instead of generating human-readable reasoning steps, the system thinks in a compact latent space \u2014 states that capture spatial information rather than producing text. The architecture alternates between two kinds of thinking: proposing candidate actions, then predicting what the world will look like if those actions are taken.\u00a0<\/span><\/p>\n It uses that predicted world state to refine its next step. It\u2019s the same reasoning loop \u2014 just in a more computationally efficient form than natural language.<\/span><\/p>\n The result: comparable output trajectory quality to text-based reasoning, using roughly half the tokens.\u00a0<\/span><\/p>\n The model was built on <\/span>NVIDIA Alpamayo<\/span><\/a> and trained using supervision derived from existing vehicle data.<\/span><\/p>\n\n
The First Foundation Model for Grasping<\/b><\/h2>\n
Teaching Autonomous Vehicles to Think Faster<\/b><\/h2>\n