Nvidia used CES 2026 to lay out an expansive bet on what it calls physical AI: robots and machines that can perceive, reason, and act in the real world without being confined to single-purpose, preprogrammed tasks. The company announced a full-stack package of robot foundation models, simulation software, workflow tooling, and new edge hardware—moves that collectively signal an ambition to become a default platform for generalist robotics, in the way mobile operating systems became foundational to the smartphone era.
The announcements reflect an industry shift as AI capabilities move beyond cloud-hosted chatbots and into embodied systems. Cheaper sensors, improved simulation, and more generalizable models are making it practical to train robots in virtual environments and deploy them on-device, where latency, reliability, and privacy constraints often make cloud dependence impractical.
A new lineup of robot foundation models
At the center of the push is a set of open foundation models released by Nvidia and made available through Hugging Face. The company framed the models as building blocks for robots that can reason, plan, and adapt across changing environments—an attempt to move beyond narrowly trained bots that perform one task well but struggle when conditions vary.
Cosmos world models and robot evaluation
The company introduced Cosmos Transfer 2.5 and Cosmos Predict 2.5, described as “world models” designed to support synthetic data generation and robot policy evaluation in simulation. In practice, world models can help developers create richer training environments and test how well a robot’s learned behavior holds up before deploying to real hardware, reducing the number of costly, slow, or risky physical trials.
Cosmos Reason 2 for vision-language reasoning
Cosmos Reason 2 was positioned as a reasoning vision-language model (VLM) aimed at helping systems “see, understand, and act” in physical settings. Vision-language reasoning has become a key frontier for robotics because it can connect natural-language instructions with visual perception and action planning—enabling robots to interpret goals rather than follow rigid scripts.
Isaac GR00T N1.6 targets humanoids
For humanoid development, Nvidia unveiled Isaac GR00T N1.6, a vision-language-action (VLA) model purpose-built for humanoid robots. The company said GR00T relies on Cosmos Reason as its “brain” and is designed to unlock whole-body control, allowing humanoids to move while manipulating objects—an important capability for tasks that require coordinated locomotion and dexterity.
Simulation and standards: Isaac Lab-Arena
Alongside the models, Nvidia announced Isaac Lab-Arena, an open source simulation framework hosted on GitHub. The pitch: as robots take on more complex work—from precise object handling to tasks such as cable installation—validating performance in the physical world becomes expensive and time-consuming, and failures can damage equipment or create safety hazards.
Isaac Lab-Arena aims to make virtual testing more systematic by consolidating task scenarios, training tools, and benchmarks. Nvidia highlighted established benchmarks including Libero, RoboCasa, and RoboTwin, positioning the framework as a step toward a unified standard in an ecosystem that has often lacked common measurement and repeatable evaluation.
OSMO as connective workflow infrastructure
To tie the workflow together, Nvidia also emphasized Nvidia OSMO, described as an open source “command center” that integrates the pipeline from data generation through training across desktop and cloud environments. For robotics teams, the operational complexity of managing datasets, simulation runs, training jobs, and deployment targets can be as challenging as model development itself. By offering orchestration and integration tooling, Nvidia is attempting to make its ecosystem not only powerful but also convenient—an important lever in becoming a default platform.
Edge compute: Jetson T4000 with Blackwell
On the hardware side, Nvidia introduced a new Blackwell-powered Jetson T4000 graphics card, the newest member of its Thor family. The company described it as a cost-effective on-device compute option for robotics developers, advertising 1200 teraflops of AI compute and 64GB of memory at an efficient 40–70 watts power envelope.
The focus on edge efficiency underscores the broader direction of robotics: many applications—industrial automation, field robotics, and mobile humanoids—require local inference and control loops that cannot rely on consistent bandwidth or tolerate cloud latency.
Deepening ties with Hugging Face and LeRobot
Nvidia also expanded its partnership with Hugging Face to broaden access to robot training and experimentation. The collaboration integrates Nvidia’s Isaac and GR00T technologies into Hugging Face’s LeRobot framework, linking what the companies describe as 2 million robotics developers in Nvidia’s ecosystem with 13 million AI builders on Hugging Face.
As part of the developer push, the open source Reachy 2 humanoid platform is now compatible with Nvidia’s Jetson Thor chip, enabling experimentation with different models without forcing developers into a single proprietary stack.
Why Nvidia is making an “Android for robotics” play
The strategy is clear: if Nvidia can make robotics development easier—by providing models, simulation, orchestration, and edge compute in a cohesive package—it can become the underlying infrastructure layer for a fast-growing market. That would place the company in a powerful position as robotics moves from pilots and labs into mainstream industrial, commercial, and eventually consumer deployments.
Nvidia pointed to early momentum: robotics has become the fastest-growing category on Hugging Face, with the company’s models leading downloads. It also said a range of robotics and industrial firms—including Boston Dynamics, Caterpillar, Franka Robots, and NEURA Robotics—are already using its technologies.
Whether the company can translate that traction into a durable standard will depend on how well its open tools interoperate with competing platforms, how quickly developers can ship reliable real-world systems, and how effectively Nvidia can balance openness with its commercial hardware advantages. But CES 2026 made one point explicit: Nvidia wants to be the default operating layer for the next generation of robots.
Related
