Runway moves into robotics with open-weight Praxis-1 AI model. Discover more Source Suppliers Rent Machinery Build Robots Runway, one of the best-known developers of generative AI video technology, is moving into robotics with Praxis-1, an open-weight AI model designed to use knowledge learned from video to control physical robots. "Open-weight" means developers can download the model and run or adapt it themselves, rather than only accessing it through Runway's servers. The company describes Praxis-1 as its first "world action model" and says it is based on the same large-scale video pretraining underlying its general world models. Runway is currently testing Praxis-1 with robotics companies Noble Machines, Standard Bots and Ultra Robotics on several types of robot hardware. It plans to make the model publicly available in the coming months. The move takes Runway beyond its established generative video business and into the rapidly developing physical AI and robotics market. Discover more Automation Consulting Service Industrial Robot Sales AI Software Development Runway says: "Today Robotics & Automation News is announcing Praxis-1, its first open-weight world action model. It's built on the same large-scale video pretraining behind its general world models. "We're actively testing Praxis-1 with early partners across a variety of embodiments, and will release it publicly in the coming months." Using internet video to train robots. The central idea behind Praxis-1 is that robots do not necessarily need to learn predominantly from enormous quantities of robot-generated training data. Collecting real-world robotics data is expensive and time-consuming, particularly when it requires robots to be teleoperated through tasks repeatedly. Runway instead wants to exploit the vastly larger quantities of ordinary video available for training AI models. Video can teach a model about object behavior, movement, physical interactions and how humans perform tasks. Runway's argument is that a robot policy beginning with this understanding of the physical world has an advantage over one trained primarily from robot action data. The company says: "A policy that already understands physical plausibility and object behavior from video pretraining has an enormous head start on one built from action data alone." Runway compares the approach with the development of large language models, which acquire a broad understanding of language from large quantities of text before being adapted for more specific tasks. The company says its experiments indicate a similar scaling effect in robotics, with robot policy performance improving as the amount of third-person video used during training increases. Web video approaches robot training data. One experiment presented by Runway compares policies pretrained using ordinary web video with those pretrained using teleoperated robot video. After fine-tuning, the company reports a final placement error of 16.1 cm for the model initially trained on web video, compared with 16.0 cm for the model trained using teleoperated robot footage. Runway notes that the difference is not statistically significant within the reported uncertainty. The company is also testing Praxis-1 on more difficult manipulation problems involving cluttered environments, transparent objects, deformable materials and groups of similar-looking objects. In another demonstration, Runway shows the same policy operating in both a controlled studio environment and a domestic kitchen without retraining. Runway brings world models into robotics. Praxis-1 builds on Runway's development of interactive video and world models including Solaris and GWM Worlds 2. These models attempt to represent how environments and objects change over time rather than simply generating individual images or predetermined video sequences. Runway says: "By teaching our models how to generate accurate physics - how objects behave, how hands move, what a task looks like partway through - we've created dynamic, complex environments for agent training in the digital and physical world." Praxis-1 applies that approach to robotics, with Runway aiming to develop a general-purpose policy capable of working across different robot embodiments and environments. The company also says previous experiments found that simulating robot policies inside its world model produced results with a 0.95 correlation to subsequent real-world performance. Testing on three different robot platforms. Praxis-1 is being tested with three early robotics partners. Noble Machines is evaluating the model for bimanual manipulation, while Standard Bots is running it on its RO1 six-degree-of-freedom robotic arm. Ultra is testing Praxis-1 with a mobile robotic platform. Runway says the early-access program will be expanded to additional companies ahead of the model's public release. The testing program is intended to evaluate both performance and safety across different robot configurations and operating environments. Praxis-1 will be released with open weights. Runway also plans to make Praxis-1 available with open weights rather than restricting access to a proprietary hosted model. The company links that decision partly to its view that greater interoperability between AI models and robot hardware will be important as physical AI develops. Runway says: "We view open world models as a compounding advantage that gives hardware developers flexibility and control they don't currently have." The company is currently accepting applications from robotics developers interested in testing Praxis-1 on their hardware before its wider release. The launch puts Runway into an increasingly important area of AI research: using large-scale world models and knowledge acquired from video to train machines capable of interacting with the physical world. If the approach scales, the significance of Praxis-1 may lie in reducing one of robotics' persistent bottlenecks - obtaining enough diverse real-world training data to produce robots that can operate reliably outside tightly controlled environments.

Runway's Praxis-1 teaches robots physical skills using web video. Runway is bringing its video pretraining stack to robotics with an open-weight policy model that generalizes across robot bodies and environments. · 13 hrs ago Read 5 min * Runway announced Praxis-1, an open-weight world action model for robotics built on its video pretraining stack. * Claim: web video pretraining matches teleop-video pretraining on final placement error (16.1 vs 16.0 cm). * One policy runs across embodiments including bimanual rigs, 6-DoF arms, and mobile bases without retraining. * Targets known failure cases: transparent objects, cluttered scenes, deformables, and repeated near-identical items. * Early partners include Noble Machines, Standard Bots, and Ultra, each running Praxis-1 on their hardware. * Weights will be released publicly in the coming months; early access is open via Runway's robotics team. Runway brings video pretraining to robot control with Praxis-1. Runway is applying its video-generation research to physical robots. The company announced Praxis-1, its first open-weight world action model, built on the large-scale video pretraining used for its general world models. Early partners are testing the model on their hardware. Runway says the weights will arrive in the coming months, although it has not provided a release date. Praxis-1 addresses one of robotics' persistent constraints: collecting demonstrations through teleoperation is slow and expensive. Runway proposes using web video to teach a model about motion, object interactions, and physical plausibility before fine-tuning it on robot-specific data. The approach could reduce the volume of demonstrations required for each task and hardware configuration. Web video targets the data bottleneck. General-purpose robot policies need examples covering different objects, environments, viewpoints, and failures. Teleoperation can capture those examples with action labels, but collecting enough data for rare conditions becomes costly. Web video offers broader visual coverage, even though it usually lacks the commands, joint positions, and force measurements recorded by robots. Runway reports that performance improves as it increases third-person video pretraining. In one placement experiment, web-video pretraining and teleoperated robot-video pretraining produced nearly identical final errors after fine-tuning: Runway's reported placement results | Pretraining source | Final placement error | | Web video | 16.1 cm | | Teleoperated robot video | 16.0 cm | The 0.1 cm difference indicates parity within this reported experiment. Broader conclusions require task definitions, dataset sizes, variance across runs, and results from independent evaluators. Those details are absent from the announcement. One model predicts and acts. A world model predicts how a scene may change over time, sometimes in response to an action. A policy model maps observations from cameras and sensors to commands a robot can execute. A world action model combines those functions, using learned representations of physical behavior to select actions. Praxis-1 builds on Runway's interactive video systems, including Solaris and GWM Worlds 2. Those systems generate controllable scenes intended to preserve physical relationships between objects, movement, and the surrounding environment. Action-free video still leaves a grounding problem. Pixels can show a hand lifting a cup, but they do not provide the joint angles, gripper force, or control frequency required to reproduce the motion. Praxis-1 therefore needs robot-specific fine-tuning to connect visual concepts with executable commands. Runway has not disclosed how that grounding works or how much labeled robot data it requires. Runway also reports a 0.95 correlation between policy evaluations performed inside its generated simulations and results on physical robots. A reliable relationship could let developers train and screen policies without repeatedly occupying a physical rig. The reported correlation does not establish absolute accuracy, calibration on rare failures, or safety under distribution shifts. Four stubborn manipulation problems. Runway organizes the model's target failure modes into four categories commonly associated with imitation-learning systems: | Category | Challenge for a robot policy | | Rigid and repeated | Distinguishing one target among many nearly identical objects | | Cluttered | Reasoning through occlusion and overlapping objects | | Transparent | Estimating shape and depth when visual cues are weak | | Deformable | Handling cloth and other objects without fixed grasp points | Runway's hypothesis is that broad video pretraining supplies useful priors about refraction, folding, object boundaries, and plausible motion. Demonstration-only policies may encounter too few examples to learn those properties reliably. The announcement does not include per-category success rates or comparisons with public baselines. Partners test the first builds. Runway is distributing Praxis-1 to selected partners before releasing the weights. The initial deployments cover several embodiments, meaning different combinations of robot bodies, arms, sensors, and control systems. | Partner | Hardware configuration | | Noble Machines | Bimanual manipulation system | | Standard Bots | RO1 six-degree-of-freedom arm | | Ultra | Mobile robot base | Runway also demonstrates the same policy operating in a controlled studio and a domestic kitchen without additional training. That test targets cross-environment transfer, which often requires site-specific fine-tuning or extensive domain randomization. The partner runs remain early product tests, with no independent benchmark results available. Open weights, unresolved license. Runway says Praxis-1 will ship with downloadable model weights. Open-weight access lets developers inspect, fine-tune, and host model parameters under the eventual license. It does not define access to training code, datasets, architecture details, or unrestricted commercial use. A permissive release could give robotics teams a general-purpose starting point for adapting policies to specific arms and mobile platforms. Teams would still need calibration data, task demonstrations, safety testing, and an integration layer for their sensors and controllers. Runway also presents open world models as part of its strategy for supporting United States development in physical AI. What developers still need. Runway's current announcement leaves several implementation and evaluation questions unanswered: * Model size, architecture, checkpoint format, and numerical precision * Inference hardware, latency, memory use, and supported control rates * Observation formats, action representations, and sensor requirements * The composition and provenance of the video pretraining corpus * The method used to ground action-free video in robot commands * The amount of robot data required for fine-tuning a new embodiment * Supported arms, grippers, mobile bases, and calibration procedures * Benchmark comparisons with Pi-0, RT-2, Octo, and other public policies * Safety constraints, failure detection, and recovery behavior * License terms, commercial rights, and a specific release date Release and replication come next. Praxis-1 presents a testable claim that large-scale video pretraining can reduce robotics' dependence on teleoperated demonstrations. The reported placement result supports that direction within one experiment, while the missing paper, weights, and benchmark details prevent reproducible comparison. Robot data will remain necessary for action grounding, hardware adaptation, evaluation, and safety. If independent tests confirm Runway's scaling results, teams may be able to use broad video pretraining to reduce the amount of teleoperation required for each deployment. Developers can request early access from Runway's robotics team.