What Runway Builds

Runway builds foundation software and world action models that control robotic systems. The company does not manufacture its own robotic arms, mobile platforms, or humanoids. Instead, Runway provides physical artificial intelligence software to robotics manufacturers, corporate research labs, and system integrators who need policy control models for existing hardware. Its core robotics offering centers on Praxis-1, an open-weight world action model that converts third-person video data into multi-axis robotic motion commands. Runway couples this with general world models, such as GWM-1, and simulation platforms that allow teams to test physical trajectories in synthetic digital environments before deploying them to physical machines. Runway claims these world model simulations predict real-world results with a 0.95 correlation factor, which reduces the need for expensive motion capture rigs and laser scanning environments. The commercial model blends free open-weight model downloads for research with enterprise licensing, application programming interface access, and tailored partner agreements for commercial deployments. Runway targets engineering teams that want to bypass manual teleoperation data collection by training robot motor skills directly on massive video repositories.

Runway Robots and Models

These are the Runway models a buyer is most likely to be evaluating, and what each one actually does.

  • Praxis-1 World Action Model — Praxis-1 is an open-weight foundation model that translates video observations into motor actions for diverse robotic hardware. The software trains primarily on third-person video rather than teleoperated demonstrations, allowing developers to expand robot control policies using web-scale video datasets. Operators load the model weights onto local compute hardware to govern end-effector trajectories, joint angles, and spatial interactions across varied real-world environments without requiring direct manual teleoperation runs.
  • GWM-1 General World Model — GWM-1 is a generative world model designed to simulate physical environments and forecast future visual states for autonomous systems. The model predicts object dynamics, fluid transitions, and mechanical interactions, giving developers an interactive environment to stress-test motion policies. Robotics teams run GWM-1 to simulate manipulation failures in software before deploying physical motor commands to expensive robotic arms or mobile bases.
  • Interactive Simulation Engines (Solaris and GWM Worlds 2) — Solaris and GWM Worlds 2 are interactive, real-time video generation systems that serve as dynamic training grounds for artificial intelligence agents. These systems generate physically consistent scenes showing hand movements and progressive object deformations. By feeding synthetic physics scenarios to reinforcement learning frameworks, the software allows robotics engineers to train manipulation policies across thousands of edge cases without causing mechanical wear on physical hardware.

Runway Pricing and Availability

Runway has not published fixed commercial price lists for Praxis-1 deployment tiers. The base architecture uses an open-weight distribution model, allowing developers to download the core neural network weights directly for local evaluation and research without upfront software licensing fees. However, production deployment costs depend heavily on the compute hardware required to run inference at acceptable control frequencies. Deploying Praxis-1 in commercial manufacturing or logistics operations requires high-performance graphics processing units (GPUs) such as NVIDIA Jetson Orin modules for onboard edge processing or dedicated data center servers for offloaded computation. Organizations running private fine-tuning pipelines must budget for custom training compute, data labeling, and specialized engineering labor to adapt the generalized model weights to their specific robot kinematic configurations. Commercial access tiers, enterprise fine-tuning support, and proprietary training environment licenses require direct engagement with Runway sales teams. Model weights are scheduled for phased public release following partner evaluations, with global availability accessible across North America, Europe, and Asia through Runway infrastructure and standard developer repositories.

Who Runway Robots Are For

Runway software suits hardware manufacturers, robotics software startups, and enterprise research teams that already operate physical robot platforms but face bottlenecks collecting real-world demonstration data. Engineering teams working with dual-arm manipulators, mobile manipulators, and humanoid platforms can utilize Praxis-1 to test video-based policy generalization across unstructured tasks like sorting, bin picking, and light assembly. Runway is not suited for end-user facilities seeking turnkey factory automation out of the box. A machine shop looking to automate machine tending or a warehouse operator requiring an off-the-shelf automated guided vehicle should look elsewhere. Praxis-1 is an underlying foundation model, not a packaged robot with an industrial controller, functional safety enclosure, or programmable logic controller interface. Organizations without dedicated machine learning personnel, custom sensor integration pipelines, and access to modern GPU hardware should deploy pre-configured commercial robots from industrial manufacturers rather than developing custom policy pipelines around Runway foundation models.

  • Verify onboard inference latency on targeted edge hardware like NVIDIA Jetson Orin modules to confirm control frequencies remain fast enough to maintain physical stability.
  • Evaluate physical safety certification pathways independently, because open-weight neural policies cannot provide deterministic motion guarantees required under ISO 10218-1:2025 and ISO/TS 15066 standards.
  • Calculate the total cost of inference infrastructure, including edge GPUs, power delivery, thermal management, and host system memory required alongside standard robot controllers.
  • Assess software integration requirements between Runway Python-based policy outputs and real-time industrial communication protocols like EtherCAT, ROS 2, or proprietary vendor controllers.
  • Review policy fine-tuning requirements to determine how much domain-specific teleoperation or physical validation data your facility must collect to achieve commercial reliability targets.
  • Determine long-term maintenance costs and model update support, establishing how custom fine-tuned weights will be managed when upstream base architectures receive revisions.

Runway Alternatives

Robots and vendors a buyer typically cross-shops against Runway.

  • Google DeepMind — Google DeepMind develops robotics foundation architectures such as RT-2 and Gemini-based control models. Unlike the open-weight release planned for Praxis-1, Google DeepMind models generally remain proprietary cloud-hosted services or research prototypes requiring integration through Google cloud infrastructure.
  • Covariant — Covariant builds the RFM-1 foundation model tailored specifically for warehouse robotics and logistics picking. While Runway targets general video pretraining across diverse embodiments, Covariant packages its models into production commercial workcells with turnkey industrial manipulation software.
  • Figure AI — Figure AI designs end-to-end humanoid hardware and neural network control systems for commercial manufacturing. Figure pairs proprietary vision-language-action policies directly with its own humanoid chassis, whereas Runway delivers hardware-agnostic software intended for third-party robotics platforms.
  • Physical Intelligence — Physical Intelligence develops generalist robot foundation policies, including the π0 model, focused on dexterous physical manipulation across commercial robotic arms. Their architecture focuses squarely on multi-task tabletop manipulation, competing directly with Runway video-derived world action models.

Recent news

A dated log of developments, each tied to its primary source.

  • 2026-10 — Runway introduced Praxis-1, an upcoming open-weight world action model that leverages large-scale video pretraining to generate robot control policies. The company is currently testing the model across multiple embodiments and environments with early partners, with plans for a public release in the coming months. The Robot Report
  • Public release dates and license terms for Praxis-1 open weights on developer repositories like Hugging Face and GitHub.
  • Published benchmark data comparing Praxis-1 real-world task completion rates against traditional teleoperated imitation learning policies.
  • Commercial partnerships with major robot arm manufacturers or humanoid developers adopting Runway software for factory pilot deployments.

Bottom Line

Runway offers a software-centric path toward general robot intelligence by substituting expensive physical teleoperation datasets with scalable video pretraining. While the Praxis-1 open-weight architecture provides significant flexibility for robotics researchers and hardware developers, it requires substantial in-house machine learning engineering, edge compute integration, and external safety infrastructure to deploy safely. Industrial operators needing predictable automation should pursue certified industrial platforms, while advanced robotics labs should evaluate Praxis-1 once public weights and hardware benchmarks release.

Explore our complete robotics database to compare hardware platforms, autonomous mobile robots, and artificial intelligence foundation models.

FAQs

Does Runway manufacture physical robot hardware?

Runway does not manufacture physical robot hardware. The company develops software architectures, foundation models, and video generation tools that run on computers and embedded processors. Hardware manufacturers, academic laboratories, and automation integrators purchase or download Runway software models to control their own mechanical robot arms, mobile bases, and humanoid platforms.

What is the primary function of Runway Praxis-1?

Praxis-1 translates visual information and task instructions into direct motor control actions for physical robots. Instead of relying solely on physical robot demonstrations collected via teleoperation, the model learns physical interactions and object dynamics from large volumes of video footage, establishing generalizable manipulation policies across varied physical environments.

How can developers access Runway robotics models?

Developers will access Praxis-1 through planned open-weight releases on standard machine learning repositories like Hugging Face and GitHub following private partner testing. Runway also provides proprietary access to broader video and world model engines through enterprise licensing agreements and cloud application programming interfaces managed directly by the company.

Can Praxis-1 replace a traditional industrial robot controller?

Praxis-1 cannot entirely replace a deterministic industrial robot controller. Industrial environments require certified programmable logic controllers and safety relays to meet regulatory standards like ISO 10218. Praxis-1 operates as a high-level trajectory and policy planner that sends path commands to low-level motor controllers responsible for deterministic joint positioning.

What compute hardware is required to run Praxis-1?

Running Praxis-1 requires modern graphics processing units capable of executing large transformer neural networks with minimal latency. Onboard edge implementations generally require compute modules like the NVIDIA Jetson Orin series, while centralized testing environments use high-performance data center GPUs to achieve the frame rates necessary for closed-loop visual control.

How does Runway train robot policies using video data?

Runway trains its models on massive collections of third-person video demonstrating physical interactions, tool usage, and environmental dynamics. By learning physical causality, object permanence, and task progression from visual feeds, the underlying neural network understands how physical tasks unfold without demanding thousands of hours of manual robot joystick demonstrations.

Primary Sources