How world model companies won U.S. robotics teams in 18 months

A world model launch with three robot companies on the press release

Fact On October 2, 2026, Runway, best known for video tools used by filmmakers, introduced Praxis-1, an open-weight world action model meant to control robots across different bodies and settings. The announcement named three partners already testing it: Noble Machines, Standard Bots and Ultra. Runway said it would publish the weights openly in the coming months, and its CTO Kamil Sindi framed the release as something that "requires open American models."7

Fact Runway was not early. Google DeepMind opened Genie 3 to a small research group in August 2025, NVIDIA had put Cosmos in front of robot and self-driving companies at CES that January, and World Labs shipped Marble to the public in November 2025. By late September 2026 AMD had agreed to buy World Labs for about $8.2 billion in stock, and Fei-Fei Li was set to become AMD's chief scientist.181216

Our read For a robot company, these launches decide which simulation and data tools its engineers will reach for next year. For a model company arriving in the U.S., they are a template. Across six companies the same four moves recur, and the companies that skipped one paid for it in slower adoption by robotics teams.

Gated previews came before public access

Fact DeepMind released Genie 3 on August 5, 2025 as a limited research preview for a small group of academics and creators. The model generated worlds at 720p and 24 frames per second that held together for a few minutes. Nearly six months later, on January 29, 2026, Google opened a prototype called Project Genie to U.S. Google AI Ultra subscribers aged 18 and over, capped at 60 seconds per generation and labeled an experimental research prototype.12

Fact The others followed similar steps. World Labs previewed Marble, kept it with early beta users for about two months, then made it available to everyone on November 12, 2025. Runway's GWM Robotics shipped as a Python SDK behind a request-access form. Odyssey unveiled Odyssey-3 in September 2026 and then opened a free browser research preview, with API access by application.12417

Our read The gate does two jobs. It keeps an unstable model away from reviewers who would judge it as a finished product, and it produces a short list of serious users whose feedback and logos are worth more than raw sign-ups. Google's half-year wait drew little backlash. A model company that opens fully on day one gives up both the feedback loop and the chance to choose who speaks first.

Named robot partners carried the credibility

Fact At CES in January 2025, NVIDIA said Waabi, Wayve, Foretellix and Uber had committed to piloting Cosmos. By GTC in March 2026 the list ran across humanoid makers including 1X, Agility, Boston Dynamics and Figure using Cosmos with Isaac Sim and Isaac Lab, robot-brain developers Skild AI and FieldAI using it for data generation, and CMR Surgical training its Versius system on Cosmos-H simulation.89

Fact Google got its strongest physical AI reference from inside Alphabet. On February 6, 2026, Waymo said its new Waymo World Model is built on Genie 3 and produces both camera and lidar data, letting engineers simulate rare events such as a tornado or an elephant in the road. Runway said in February 2026 that it was building GWM Robotics with NVIDIA and Berkshire Grey.36

Fact The smaller companies did the same at smaller scale. World Labs published a case study the day Marble launched showing Lightwheel turning a single 360-degree photo into a robot test environment inside NVIDIA Isaac Sim. Odyssey's launch featured Flexion, a humanoid company that built control policies on Odyssey-3.1318

Our read Robotics buyers are a small, connected group, and they read partner lists closely. A logo means a team with hardware and a deadline has put engineering hours into the model. TechCrunch's GWM-1 launch story could only report that Runway was in talks with unnamed robotics firms; by Praxis-1 ten months later, Runway had three named partners. That gap is what the first months should close.5

Open weights fed a builder community

Fact NVIDIA released Cosmos in January 2025 under its permissive open model license, which allows commercial use, after training on about 20 million hours of real-world video. TechCrunch noted the models were not open source in the strict sense, because NVIDIA did not publish full training data details. In March 2026 NVIDIA and Hugging Face integrated Isaac and GR00T into the open LeRobot framework, linking what NVIDIA called 2 million robotics developers with Hugging Face's 13 million AI builders.89

Fact NVIDIA backed the release with cheap community programs. The Cosmos Cookoff, a virtual hackathon that ran from late January to February 26, 2026, offered $5,000 in total cash, with a DGX Spark for first place and judges from Hugging Face, Nebius and Nexar. When Cosmos 3 arrived on May 31, 2026, NVIDIA listed Runway, Skild AI, Generalist and Agile Robots among founding members of a Cosmos Coalition.1110

Fact Tencent used open releases to claim firsts. Its HY-World 1.5, released December 17, 2025, called itself the first open-source, real-time interactive world model with long-term geometric consistency, and Tencent released training code along with the weights; a 5B version followed on January 6, 2026. HY-World 2.0, released April 16, 2026, exports meshes and Gaussian splats that Tencent says can go into game engines and embodied simulation pipelines.2021

Our read Open weights matter more in robotics than in creative video because robot teams need to fine-tune on their own sensors and run models on their own hardware. A closed API cannot sit inside a training loop on a factory floor. Runway's move from a gated SDK to an open-weight action model, with explicitly national framing, shows that even a closed-model company concluded it needed weights to win this buyer.

Showcases sorted by job rather than by model

Fact World Labs' Marble Labs groups its case studies into three themes: Imagine for film, VFX and virtual production; Explore for gaming and immersive media; and Simulate for robotics, architecture and health systems. Runway split GWM-1 into three separately trained models, GWM Worlds, GWM Avatars and GWM Robotics, and described robotics uses as synthetic data augmentation and policy evaluation in simulation.144

Fact World Labs kept its biggest commercial bet on the creative side. Its $1 billion round in February 2026 included $200 million from Autodesk, and TechCrunch reported that the partnership would start with media and entertainment uses, while robotics stayed one tab among several on Marble Labs.1514

Our read A robotics engineer arriving at a model company's site wants to know one thing: does this help me collect less data, or test policies faster? Separating the robotics page from the filmmaker gallery answers that within a minute. Marble's "Simulate" tab and Runway's named use cases both do this. A model company that leads with cinematic reels will be filed under media and will not get a meeting with a robot team.

Evidence a robotics engineer can check

Fact Runway's February 2026 study ran eight manipulation policies from the RoboArena benchmark on a Franka arm inside GWM Robotics, with human graders scoring 1,450 simulated rollouts. Rankings in simulation matched real-world rankings with a Pearson correlation of 0.95, the company reported. Odyssey said a driving policy trained on about 20 hours of model-generated data drove on real streets in India.619

Fact Lightwheel's case study claimed environment creation dropped "from weeks to minutes." These figures are company-reported. The Robot Report presented Runway's 0.95 as Runway's claim, and Tech Times described the India drive as a limited real-world test rather than a commercial deployment.13719

Our read Self-reported numbers still work if they are specific and reproducible: a public benchmark, a named arm, a stated number of rollouts. That is the format robotics researchers use themselves. A model company should publish its first robotics result in that format, on a benchmark others can rerun, before it publishes anything else about physical AI.

A 90-day plan for a model company entering U.S. physical AI

Our read Days 1 to 30: pick one robotics job and one benchmark. Choose policy evaluation, synthetic data for manipulation, or environment generation, not all three. Recruit a cohort of 10 to 20 U.S. robotics teams, the way Genie 3 started with a small group, and give them gated access in exchange for a joint result. Build the integration robot teams already use, such as an export into Isaac Sim or a LeRobot-compatible checkpoint, before building a consumer demo.

Our read Days 31 to 60: turn the cohort into named partners. Aim for two or three companies willing to appear on the launch release, as Runway did with Noble Machines, Standard Bots and Ultra. Publish a reproducible result in the format of Runway's RoboArena study. Release a small open-weight model, on the pattern of Tencent's 5B WorldPlay or NVIDIA's Cosmos license, so builders can fine-tune without a sales call.

Our read Days 61 to 90: launch where robotics buyers already gather, and keep the community cheap. A four-week virtual hackathon with hardware prizes cost NVIDIA $5,000 in cash. Split the website by job, with a robotics page that shows the benchmark, the partners and the SDK. Companies arriving from outside the U.S. should expect questions about data origin and hosting, since Runway has already framed open American models as a selling point, and should have written answers ready before launch day.