NVIDIA Physical AI: Cosmos Reason 2, GR00T N1.6, and the Robots Coming to Work

NVIDIA just drew the clearest map yet of how physical AI will be built. During National Robotics Week 2026, the company released a coordinated set of open models, infrastructure tools, and hardware that together form what NVIDIA calls the physical AI stack — the software and silicon layer that will power the next generation of intelligent machines. The centerpieces are Cosmos Reason 2 and Isaac GR00T N1.6, but the full announcement spans world models, evaluation frameworks, edge compute hardware, and a growing roster of robot partners bringing these models into production. Here is every piece, what it does, and why it matters.

What NVIDIA Means by Physical AI

When NVIDIA talks about “physical AI,” it means AI systems that perceive the physical world through sensors, reason about what they observe, and take actions through physical actuators — robots, autonomous vehicles, industrial machines. This is fundamentally different from language AI, where the system’s actions are tokens in a text stream. Physical AI must deal with continuous sensor data, real-time motion constraints, the irreversibility of physical action, and the near-infinite variability of real environments.

The challenge of physical AI is not perception alone — cameras and sensors have been capable enough for structured environments for years. The challenge is reasoning and generalization: a robot needs to not just recognize a cup on a table but understand what to do with it in the context of a task, handle unexpected variations in the environment, and execute the appropriate physical motion with the right dynamics. This is the capability gap that NVIDIA’s new model family is designed to close, and the National Robotics Week release is the most coherent public demonstration of that ambition to date.

Cosmos Reason 2: The Vision-Language Brain for Robots

The foundation of the new physical AI stack is Cosmos Reason 2, an open reasoning vision-language model (VLM) that enables intelligent machines to see, understand, and reason about the physical world. Unlike general-purpose VLMs built primarily for text and image tasks, Cosmos Reason 2 is trained and post-trained specifically on data relevant to physical environments — spatial relationships, object properties, task sequences, and the cause-and-effect logic of physical actions.

Cosmos Reason 2 is available in a 2-billion-parameter variant called Cosmos-Reason-2B, which is small enough to run on edge hardware while powerful enough to deliver meaningful scene understanding and task decomposition. The model uses native resolution image input — a technical choice that matters more than it might appear. Many smaller VLMs resize or crop input images before processing, which introduces distortion and information loss. Native resolution processing means the model sees what the robot’s cameras actually capture, without artifacts that degrade spatial reasoning.

Cosmos Reason 2 is post-trained on specialized physical AI data using NVIDIA Isaac Sim pipelines, enabling the model to generalize across millions of simulated scenarios without requiring a distinct training run for every environment variation. For robot developers, this is the difference between a system that works in the lab and one that can be deployed in production — generalization to the unexpected is the single hardest property to engineer into a robotic system, and simulation-based post-training is the most scalable path to achieving it.

Isaac GR00T N1.6: Full-Body Humanoid Robot Control

Built directly on top of Cosmos Reason 2, Isaac GR00T N1.6 is NVIDIA’s most advanced open humanoid robot model. GR00T N1.6 is a vision-language-action (VLA) model — meaning it takes visual input, processes natural language instructions, and outputs physical actions — specifically designed for full-body humanoid control. It is the combination of perception, reasoning, and motion that prior systems handled in separate, loosely coupled components, unified into a single end-to-end trainable policy.

Architecture: A 32-Layer Diffusion Transformer

GR00T N1.6’s action generation is handled by a 32-layer diffusion transformer. Diffusion models generate outputs by iteratively refining from noise to a coherent result, which makes them well-suited for physical motion planning: rather than predicting a single action deterministically, the model refines a motion trajectory over multiple steps, producing smoother and more physically consistent movements than direct policy outputs. For humanoid robots, which must coordinate dozens of joints with precise timing, this matters significantly in practice.

The model integrates Cosmos-Reason-2B with native resolution vision, which directly addresses one of the persistent failure modes in robot policies: poor visual grounding. When a robot misidentifies an object or misjudges spatial relationships because its vision system distorted the input, every downstream action is built on a corrupted foundation. Native resolution vision in the perception module means N1.6’s actions are grounded in accurate observations — a detail that improves reliability across the full task distribution.

Training Scale and Real-World Validation

GR00T N1.6 was pretrained for 300,000 steps with a global batch size of 16,384 — a large-scale training run that establishes broad capabilities before task-specific fine-tuning. Task-specific post-training requires only 10,000–30,000 steps with a batch size of 1,000, meaning the model can be adapted to specific robot hardware and task requirements with relatively modest compute. This sim-to-real workflow is central to NVIDIA’s strategy: train broadly in simulation at scale, then fine-tune efficiently on real-world data.

The performance improvement over the previous generation is measurable and hardware-validated. N1.6 outperforms N1.5 on simulated manipulation benchmarks and on three distinct real-world robot platforms: bimanual YAM, Agibot Genie-1, and Unitree G1. Testing across multiple hardware configurations is an important signal — it suggests that the improvements are architectural rather than specific to one robot’s quirks, and that the model generalizes across humanoid form factors. According to NVIDIA’s research documentation, the improvement in scene understanding from native-resolution Cosmos-Reason-2B directly translates to better task decomposition and more reliable manipulation outcomes.

The Supporting Infrastructure

Cosmos Reason 2 and GR00T N1.6 are models. What makes them deployable at scale is the infrastructure layer NVIDIA released alongside them.

Isaac Lab-Arena: Standardized Robot Benchmarking

Isaac Lab-Arena is a collaborative environment for large-scale robot policy evaluation and benchmarking. It integrates with established simulation benchmarks including Libero and Robocasa, providing a standardized testing environment before real-world deployment. For the physical AI field, this is the equivalent of what evaluation suites have become for language models — a common ground for measuring progress and identifying failure modes that is rigorous, reproducible, and comparable across teams. Standardized benchmarking infrastructure has historically accelerated progress in AI by making comparisons between approaches scientifically meaningful. Robots have lacked this until now.

OSMO: Edge-to-Cloud Compute for Robot Training

OSMO is NVIDIA’s edge-to-cloud compute framework designed to simplify robot training workflows. Running robot training is operationally complex — it requires coordinating between on-device data collection, cloud-based training compute, simulation environments, and real-world validation. OSMO provides a unified framework for managing this workflow, reducing the infrastructure burden that currently prevents many robotics teams from iterating quickly. The design goal is to make the robot training pipeline as manageable as modern software CI/CD pipelines are for code — automated, reproducible, and scalable without requiring a dedicated infrastructure team.

Cosmos Transfer 2.5 and Cosmos Predict 2.5

Rounding out the model releases are Cosmos Transfer 2.5 and Cosmos Predict 2.5, updates to NVIDIA’s world model family. World models in robotics are systems that can predict what the physical world will look like given a sequence of actions — they allow robots to “imagine” the consequences of actions before taking them, enabling planning without requiring physical trial and error for every scenario. Cosmos Transfer 2.5 focuses on domain transfer: bridging the gap between synthetic simulation data and real-world environments. Cosmos Predict 2.5 handles forward prediction of physical dynamics. Together, they form the world model backbone that makes sim-to-real generalization feasible at scale.

Jetson T4000: The Edge Hardware Running This Stack

All of these models require compute to run. NVIDIA’s answer for on-device deployment is the Jetson T4000, a Blackwell architecture-powered module designed for autonomous machines and robotics. The T4000 delivers 4x greater energy efficiency and AI compute than the previous generation, with 1,200 FP4 TFLOPS and 64GB of memory within a configurable 70-watt power envelope. It supports 16 lanes of MIPI CSI, enabling simultaneous ingestion from multiple camera streams — a critical capability for robots that need multi-angle environmental perception.

At $1,999 at 1,000-unit volume, the T4000 is positioned as production hardware rather than research hardware. It runs JetPack 7.1 and is designed to execute NVIDIA’s physical AI model stack locally, without cloud connectivity — which is essential for real-world robot deployment where latency and reliability requirements make cloud-dependent architectures impractical for real-time control. The Blackwell architecture delivers roughly 4x the performance per watt of the prior Jetson generation, which means models like Cosmos-Reason-2B can run in real-time on embedded hardware at the energy budgets that industrial and mobile robots operate within.

The Robot Partners Bringing Physical AI to Market

NVIDIA’s physical AI platform is designed as an ecosystem. During National Robotics Week, several major robotics companies demonstrated production deployments built on the new model stack:

• Boston Dynamics showcased autonomous machines built on NVIDIA technologies, extending the company’s hardware expertise with NVIDIA’s AI stack for behavioral control and environmental reasoning.
• NEURA Robotics used GR00T-enabled workflows for humanoid training and real-world validation, demonstrating sim-to-real transfer for bipedal manipulation tasks.
• Franka Robots integrated GR00T for manipulation tasks, demonstrating the model’s applicability to industrial robotic arms beyond humanoid platforms.
• Caterpillar unveiled autonomous construction equipment built on NVIDIA’s physical AI infrastructure, one of the highest-stakes real-world deployments for any autonomous system.
• LG Electronics demonstrated service robotics built on the platform, signaling enterprise and consumer applications beyond industrial settings.
• Salesforce is using Cosmos Reason with Agentforce and the NVIDIA Blueprint for video search to analyze footage captured by its robots, reporting a 2x improvement in incident resolution times.

The range of partners spans construction, manufacturing, consumer electronics, and enterprise software — a breadth that reflects NVIDIA’s strategy of positioning itself as the platform layer for physical AI across industries, rather than building specific robots for specific use cases.

NVIDIA’s “Android of Robotics” Strategy

NVIDIA is explicitly positioning itself as the platform standard for generalist robotics — an open, standardized foundation that multiple hardware manufacturers can build on, rather than a vertically integrated system competing directly with robot makers. The Android analogy is instructive: Google built Android not to sell smartphones but to ensure the entire smartphone ecosystem ran on Google’s platform layer. NVIDIA’s physical AI stack serves a similar function. Cosmos and GR00T are open models available on Hugging Face; Jetson is accessible hardware; the training infrastructure (OSMO, Isaac Sim) is available to partners; and NVIDIA captures value at the compute and platform layers while the ecosystem proliferates.

The openness of the models is strategically important. By releasing Cosmos Reason 2 and GR00T N1.6 under open licenses, NVIDIA creates conditions for a research and developer community to build on its platform — lowering the barrier to adoption while establishing its model family as the de facto foundation for physical AI development. The Hugging Face partnership, confirmed in the official announcement, means these models benefit from the same discoverability and integration infrastructure that has made open language models accessible to millions of developers. Based on our analysis of how the Jetson platform grew into dominance for edge AI applications, the same network effects are likely to compound in physical robotics as more developers build on this stack.

What This Means for Developers and Builders

If you are building anything that involves physical AI — robotics, autonomous vehicles, smart industrial systems, or edge AI applications — the NVIDIA National Robotics Week announcements represent the most coherent open platform publicly available today. The practical implications:

• Start with the open models. GR00T N1.6 and Cosmos Reason 2 are available on Hugging Face. For developers with access to robot hardware, these provide a significantly better starting point than training from scratch or adapting general-purpose VLMs to physical tasks. The Cosmos-Reason-2B model is small enough to run on consumer-grade hardware during development.
• Evaluate with Isaac Lab-Arena before real-world testing. Standardized benchmarking against Libero and Robocasa gives you a reproducible measurement of model performance before committing to expensive and time-consuming real-world validation.
• Plan hardware around Jetson T4000 for production. For physical AI applications requiring local inference, the T4000’s 1,200 FP4 TFLOPS at 70 watts represents the most capable edge AI module in production. At $1,999 per unit at volume, the economics work for serious commercial deployments and are well within reach for prototyping.
• Use OSMO to manage training pipeline complexity. The operational complexity of coordinating simulation, data collection, cloud training, and real-world validation is a significant engineering burden that slows iteration. OSMO is designed to reduce that burden to something a small team can manage.

The physical AI era is moving faster than most observers expected two years ago. NVIDIA’s coordinated release — open models, evaluation infrastructure, compute hardware, and an active partner ecosystem deploying in production — represents the most complete platform stack publicly available for building intelligent physical systems. According to our analysis of the robotics development landscape, teams that begin working with these tools now will be positioned ahead of those who wait for the platform to mature further, because model and infrastructure improvements are arriving in rapid succession and the compounding learning advantage grows with each iteration cycle.