NVIDIA is making a significant leap into the realm of robotics and autonomous machines with the introduction of its new physical AI models and frameworks unveiled at CES 2026. These innovations, centered around the NVIDIA Cosmos platform and the Isaac GR00T N1.6 humanoid vision-language-action (VLA) model, are designed to bridge the gap between simulation and real-world robot deployment. Accompanied by the Isaac Lab-Arena evaluation framework and the OSMO edge-to-cloud compute infrastructure, NVIDIA’s latest offerings promise to accelerate open-source robotics development, particularly through integrations with the LeRobot open-source robotics library and the popular AI community platform Hugging Face.
Physical AI and World Foundation Models: A New Robotics Paradigm
NVIDIA’s initiative revolves around the concept of physical AI — artificial intelligence systems capable of perceiving, reasoning, and interacting in the physical world. This extends beyond traditional software AI to embodied intelligence that involves perception, planning, and control running at the network edge within robots, supported by powerful cloud computing resources.
At the heart of this vision is NVIDIA Cosmos, a platform housing advanced generative world foundation models (WFMs). These models learn from and synthesize visual, temporal, and multimodal data about the physical environment, enabling realistic simulation and prediction of dynamic, real-world scenarios. With Cosmos, developers can generate synthetic video sequences for robot training, predict object and agent movements in complex settings, and perform reasoning about scenes through vision-language frameworks.
NVIDIA Cosmos: Generating and Reasoning About the Physical World
The Cosmos platform includes several key models:
- Cosmos Transfer 2.5: Focused on synthetic video generation, this model enables scalable creation of training datasets depicting robots performing various tasks in simulated environments, thus reducing reliance on costly real-world data collection.
- Cosmos Predict 2.5: Specialized in forecasting object trajectories and motion, useful for collision avoidance and dynamic navigation.
- Cosmos Reason 2: An open reasoning vision-language-action model that interprets complex multi-step instructions and understands dynamic visual contexts, effectively serving as the cognitive backbone for robot planning and decision-making.
Enterprises like Salesforce, Uber, and Hitachi already leverage Cosmos Reason to enhance productivity agents with capabilities such as live video analysis and incident response acceleration, cutting resolution times in half.
Isaac GR00T N1.6: An Open Model for Humanoids
Complementing Cosmos, NVIDIA’s Isaac GR00T N1.6 targets humanoid robots, embodying a more comprehensive intelligence that spans perception to whole-body control. This open VLA model processes egocentric camera inputs alongside natural language commands and proprioceptive data, generating nuanced loco-manipulation policies.
Isaac GR00T N1.6 features a 32-layer diffusion transformer architecture—twice the size of prior models—to produce smooth, adaptive motion. Trained on thousands of hours of teleoperation data across diverse robot types, including bimanual arms and mobile manipulators, the model generalizes across hardware, facilitating broad applicability.
A key hallmark is GR00T’s integration with Cosmos Reason 2, which enhances scene understanding and breaks down complex instructions into manageable actions within real-time contexts. This tight coupling enables humanoid robots to perform sophisticated tasks involving navigation, manipulation, and coordinated locomotion, all validated within a simulated environment before sim-to-real transfer.
Isaac Lab-Arena: Benchmarking Robotics Performance
To ensure the reliability and performance of robots powered by these models, NVIDIA introduces Isaac Lab-Arena, a standardized framework for benchmarking robotic capabilities. This evaluation environment offers predefined tasks and measurement criteria for assessing locomotion, manipulation, and navigation skills of diverse autonomous machines, including humanoids.
Isaac Lab-Arena’s deep integration with GR00T models facilitates rigorous, repeatable assessments, allowing developers to fine-tune policies and validate behavior in controlled, simulated scenarios before deploying to physical hardware. Access to this framework is simplified through the LeRobot open-source ecosystem, inviting the robotics research and development community to participate in collaborative advancements.
OSMO: Seamless Edge-to-Cloud Robotics Computation
Effective deployment of physical AI demands robust infrastructure. NVIDIA’s OSMO framework orchestrates computational tasks spanning edge devices embedded within robots to large-scale GPU clusters in data centers. This integration streamlines data collection, labeling, and synchronization across robot fleets, enabling continuous model training and version management.
OSMO manages the full lifecycle—from on-device inference and control powered by the Jetson Thor edge AI computer, to cloud-based training accelerated on NVIDIA DGX systems—and ensures updated policies are seamlessly rolled out back to robots. This operational backbone supports NVIDIA’s ambition to provide an end-to-end AI platform for robotics applications.
Open-Source Collaboration: LeRobot and Hugging Face Integration
A core aspect of NVIDIA’s strategy is fostering community-driven robotics innovation by integrating their models into established open frameworks. The LeRobot ecosystem now hosts GR00T models and Isaac Lab-Arena tools, enabling developers to easily fine-tune humanoid AI policies and conduct evaluations on accessible platforms.
Simultaneously, NVIDIA has partnered with Hugging Face to interface with open-source robots such as Reachy 2, a humanoid robot, and Reachy Mini, a tabletop robot. Reachy 2 is now fully interoperable with Jetson Thor, empowering it to run VLA models like GR00T N1.6 directly on the hardware. Reachy Mini benefits from connection to DGX Spark, enabling rich AI-driven experiences with local language processing and vision models. These integrations bridge hobbyist, academic, and industrial robotics, expanding the reach of NVIDIA’s physical AI stack.
Hardware Enablers: Jetson Thor and DGX Spark
NVIDIA’s model innovations are bolstered by purpose-built hardware. Jetson Thor equips robots with the computational power needed to execute complex VLA models in real-time, while DGX Spark provides a scalable, data-center-grade environment for training, inference, and large-scale simulation operations.
Together with OSMO’s orchestration, these platforms form a cohesive infrastructure, aligning hardware and software to enable robots that can learn robustly and adapt fluidly to varied environments.
Industry Momentum and Future Prospects
NVIDIA’s announcement has garnered partnerships from prominent robotics firms such as Boston Dynamics, Franka Robotics, Caterpillar, and NEURA Robotics. These collaborators are integrating NVIDIA’s physical AI stack into next-generation platforms for applications in industrial automation, logistics, service robotics, and autonomous machines.
Beyond robotics, enterprises including Salesforce and Hitachi employ Cosmos Reason for AI-driven enterprise agents, highlighting the framework’s versatility beyond humanoid or vehicle platforms.
This ecosystem approach signals NVIDIA’s ambition to become the “Android of generalist robotics”, providing a standardized yet flexible platform for the creation and deployment of robots capable of navigating and performing in the physical world.
Looking Ahead: Opportunities and Challenges
NVIDIA’s open models, synthetic data generation, and sim-to-real workflows offer tremendous potential to accelerate robotics R&D while reducing cost and risk. By lowering barriers to entry and uniting hardware and software ecosystems, the company is inviting a new wave of innovation in autonomous machine intelligence.
Nonetheless, challenges remain: the dependency on NVIDIA’s hardware backend raises concerns of platform lock-in among developers; the computational expense of training and deploying large VLA models can be prohibitive for smaller entities; and ensuring safety, robustness, and regulatory compliance in real-world robot operation remains an ongoing imperative.
As NVIDIA continues expanding its physical AI capabilities and community engagement, the coming years will reveal how this comprehensive robotics platform shapes the trajectory of autonomous systems across industries.
Additional Resources
