NVIDIA has expanded its robotics platform with a sweeping set of new tools, including updated GR00T open foundation models, Cosmos world models, the Newton 1.0 physics engine, and refreshed simulation software spanning Isaac Sim 6.0, Isaac Lab 3.0, and Omniverse NuRec. The announcements arrive as the robotics industry pivots from isolated hardware demos toward layered autonomy stacks combining foundation models, physics simulation, and edge compute. Alongside the robotics platform news, NVIDIA also said it can help Uber scale a global autonomous vehicle fleet toward 100,000 vehicles over time starting in 2027, a claim that, while not a confirmed deployment, signals how central NVIDIA now sees itself in both ground robotics and autonomous mobility.
***
The timing is notable. Just this week, Gravis Robotics raised $200 million for autonomous construction, Perceptron AI launched a 36-billion-parameter open-weight embodied model called Isaac 0.5, and AgiBot demonstrated robots learning new factory skills in minutes using reinforcement learning on a live production line. Taken together, these developments point to an industry-wide shift: the hard problem in robotics is no longer just building better arms, legs, or grippers, but building the simulation environments, world models, and orchestration software that let those bodies learn and operate safely at scale. NVIDIA's latest push is a direct bet that whoever controls the simulation and training layer will control the pace of that scaling.
A Full-Stack Play for Physical AI
NVIDIA's announcement bundles together several previously separate efforts into what the company is now positioning as a unified physical AI stack. GR00T, its family of open foundation models for humanoid and general-purpose robots, has been expanded alongside Cosmos, a set of world models designed to generate synthetic video and physics-consistent scenarios for training. The newly introduced Newton 1.0 physics engine is meant to close the gap between simulated and real-world robot dynamics, a persistent weak point in embodied AI development where robots trained in simulation often fail when deployed on real hardware.
Rounding out the release are updated simulation tools: Isaac Sim 6.0, Isaac Lab 3.0, and Omniverse NuRec. Isaac Sim provides the core simulated environment for testing robot behavior, Isaac Lab focuses on reinforcement learning workflows for training policies, and NuRec is aimed at reconstructing real-world scenes into simulation-ready assets. Together, the pieces suggest NVIDIA is trying to own not just the chips that run robots, but the entire pipeline that produces the software running on top of them, from data generation to training to deployment.
Why Simulation Is Becoming the Battleground
The reason simulation has become such a central focus is straightforward: real-world robot data is expensive, slow, and dangerous to collect at scale. A humanoid robot that falls while learning to walk in a factory is a liability and a delay; a humanoid robot that falls a million times in simulation costs almost nothing. Perceptron AI's Isaac 0.5, a 36-billion-parameter open-weight model combining video understanding, embodied reasoning, and robot control, and AgiBot's Real-World Reinforcement Learning system, which reportedly teaches robots new skills in minutes on a live production line, both reflect the same underlying pressure: companies need faster, cheaper ways to generate the training signal that embodied AI models require.
This is also where NVIDIA's competitive position gets more complicated. Google has been pushing Gemini Robotics ER 2 as a high-level orchestrator for robot control, publishing benchmark figures of 57.4% accuracy on progress classification and 91.3% on moment finding. UBTECH, Figure AI, Apptronik, and Agility Robotics are all fielding humanoid or wheeled robots in real pilot programs at companies like BYD, BMW, Mercedes-Benz, Schaeffler, and GXO. NVIDIA does not build robots itself, so its strategy depends on becoming the default infrastructure layer underneath all of them, similar to the role it has played in AI data centers over the past three years.
The Uber Signal and the Broader Autonomy Push
NVIDIA's statement that it can support Uber in scaling a global autonomous vehicle fleet to 100,000 vehicles over time starting in 2027 is a platform capability claim rather than a confirmed deployment, and it should be read with appropriate caution. Still, it is a meaningful marker of intent. Autonomous driving has quietly become one of the strongest commercial proof points for embodied AI generally, alongside Kodiak AI's newly announced collaboration with AMD to use EPYC series processors for driverless trucking compute. Both moves suggest that autonomy providers are locking in hardware and compute partnerships years ahead of planned scale-ups, a sign of growing confidence that regulatory and technical barriers are becoming more manageable.
The mobility and ground-robotics stories are converging in an important way. The same core technologies, world models for anticipating physical outcomes, reinforcement learning for skill acquisition, and high-fidelity simulation for safe testing, are being applied to warehouse robots, humanoids, construction equipment, and self-driving vehicles alike. That convergence is precisely why NVIDIA is pitching Cosmos and Newton as general-purpose infrastructure rather than tools built for any single robot form factor.
Standardization, Cost, and the Road to Real Deployment
Even as the technology stack matures, the industry is grappling with how to measure and standardize its real-world performance and cost. ABB Robotics is leading work with 11 other countries on an ISO Technical Specification to measure industrial-robot energy consumption and efficiency, an unglamorous but significant step that signals robotics is graduating from experimental deployment toward operational accountability. Companies buying robots at scale, whether AMRs like ABB's new Flexley Stack F712 autonomous forklift or humanoids on a factory floor, increasingly need standardized ways to compare energy use, uptime, and total cost of ownership.
That practical, cost-driven mindset is a useful check on the more speculative promises attached to humanoid robotics and simulation-heavy training pipelines. Reporting on China's humanoid ecosystem notes that despite rapid manufacturing and infrastructure growth, many systems remain too slow and error-prone for most factory work today. NVIDIA's simulation-first bet is, in effect, an argument that better training infrastructure is the fastest route to closing that performance gap, but the ISO efficiency work and continued caution from manufacturers suggest the industry is also preparing for a longer, more incremental path to widespread deployment.
The next wave of physical AI won't be won by the company with the best robot arm. It will be won by whoever can simulate a million hours of real-world failure before a robot ever touches a factory floor.
What Comes Next
NVIDIA's expanded platform will not, by itself, resolve the core challenges facing embodied AI: robots still struggle with dexterity, generalization across unpredictable environments, and safety assurance in spaces shared with humans. But by bundling GR00T, Cosmos, Newton, and its simulation suite together, the company is making a clear wager that infrastructure providers, not individual robot makers, will capture much of the value as the sector scales from pilots to production.
The coming year will test that wager directly. Hyundai's plan to build a plant capable of producing 30,000 Atlas humanoids annually by 2028, Schaeffler's scheduled scaled manufacturing of strain wave gearboxes in 2027, and Uber's prospective fleet expansion all depend on the underlying software and training systems working reliably at a scale none of these companies have yet proven. Whether NVIDIA's simulation-heavy approach becomes the industry standard, or whether rivals like Google and a wave of well-funded startups chart a different path, will likely define the competitive landscape of physical AI for the rest of the decade.
Sources
- https://www.therobotreport.com/
- https://www.roboticstomorrow.com/stream/news/
- https://www.businesswire.com/newsroom/industry/technology/robotics
- https://roboticsandautomationnews.com/
- https://www.robotics247.com/
- https://www.therobotreport.com/robotics-news/
- https://www.therobotreport.com/category/news/
- https://www.turingpost.com/p/roboticsupdates
- https://www.therobotreport.com/category/robots-platforms/industrial-robots/
- https://www.therobotreport.com/top-10-robotics-stories-of-may-2026/
- https://blogs.nvidia.com/blog/national-robotics-week-2026/
- https://www.turingpost.com/p/roboticsnews
- https://www.sciencedaily.com/news/computers_math/robotics/
Leave a Comment