Runway, the artificial intelligence company known for its video-generation technology, has released Praxis-1, a 27-billion-parameter robotics model designed to translate visual understanding into physical actions. The model is now publicly available with its weights, inference code and fine-tuning tools under the Apache 2.0 licence, allowing developers and researchers to download, modify and use it commercially. The release, announced on November 11, 2026, opens access to technology that Runway had previously tested with selected robotics partners.
Praxis-1 takes a different approach to teaching robots how to perform tasks. Instead of relying exclusively on demonstrations recorded by physical robots, it uses knowledge learned from large-scale video pretraining to predict how objects and environments will change when an action is performed. It then connects those predictions to robot movements. Runway says the model achieved a 46.2% success rate across 1,200 episodes on NVIDIA’s RoboLab-120 benchmark, placing it ahead of several competing open-weight robotics models.
What Is Runway’s Praxis-1?
Praxis-1 is a world action model designed to help robots understand their surroundings and perform physical tasks.
It combines visual prediction with action generation. Given a task instruction, camera images and information about a robot’s position, the model predicts what the environment will look like after an action and generates corresponding movements.
For example, a user could instruct a robot to pick up a tennis ball and place it inside a box. Praxis-1 processes the visual information, predicts how the ball and surrounding objects will move, and generates the actions needed to complete the task.
The model is based on a 27-billion-parameter mixture-of-experts diffusion transformer architecture. Its design uses a large video foundation model alongside a smaller action-generation component that converts learned visual representations into robot control instructions.
Runway’s central idea is that AI models can learn useful information about physical environments from ordinary videos. Such videos contain examples of how people manipulate objects, how materials behave and how everyday tasks unfold.
This could reduce the industry’s dependence on expensive robot-specific demonstration data, although adapting the model to a particular machine still requires additional training.
How Praxis-1 Uses Video to Control Robots
Traditional robotics models often depend heavily on demonstrations collected using physical robots.
Developers must record movements, label tasks and gather enough examples to teach a system how to handle different objects and environments. This process can be expensive and time-consuming, particularly when a robot needs to perform tasks in unfamiliar situations.
Praxis-1 attempts to address this limitation through video pretraining.
The model first learns patterns from large collections of video, developing an understanding of object movement, physical interactions and how tasks progress over time.
It then combines this visual knowledge with robot-specific training to generate actions.
The process can be understood in three stages:
- Video understanding: The model learns how objects, hands and environments behave from large-scale video data.
- Future prediction: It predicts how a scene may change when a particular action is performed.
- Robot control: An action-generation component converts those predictions into movements that the robot can execute.
The approach is designed to give robots a foundation of physical-world knowledge before developers fine-tune the system for a specific machine.
However, watching videos alone does not automatically teach a robot how to perform every task. Real-world deployment still requires suitable demonstrations, hardware integration and testing.
Praxis-1 Achieves 46.2% on NVIDIA RoboLab-120
One of the major claims surrounding the release is Praxis-1’s performance on NVIDIA’s RoboLab-120 benchmark.
Runway reported that the model achieved a 46.2% success rate across 1,200 evaluation episodes.
The company compared the result with several other robotics models.
| Robotics model | RoboLab-120 success rate |
|---|---|
| Runway Praxis-1 | 46.2% |
| FLUX 3 Action | 42.9% |
| Cosmos3 Nano | 36.8% |
| π0.5 | 28.0% |
<text color=”secondary” size=”xs”>Source: Runway’s November 2026 announcement. Figures are company-reported benchmark results.</text>
The results place Praxis-1 ahead of the competing models listed in Runway’s comparison.
However, these figures should be interpreted in the context of the evaluation setup. The reported RoboLab-120 result involves a model fine-tuned on robot demonstration data rather than an entirely unadapted model operating on unfamiliar hardware.
The benchmark is useful for comparing performance under specified conditions, but it does not guarantee that the same success rate will be achieved in factories, warehouses or household environments.
The results have also been published by Runway and should not be treated as independently reproduced performance measurements.
Model Reaches 80% Success After Fine-Tuning on Google’s RT-1 Robot
Runway also reported results from an experiment involving Google’s RT-1 robot.
The company fine-tuned Praxis-1 using data from RT-1, a robot platform that was not included in the model’s original pretraining.
After adaptation, Praxis-1 achieved an 80% success rate in the SimplerEnv simulation environment.
For comparison, the original RT-1 policy achieved a 52.4% success rate in the same evaluation.
This result is important because it suggests that the model’s video-based pretraining can provide a useful foundation when adapting to different robot platforms.
However, the 80% figure does not represent zero-shot performance. Praxis-1 was fine-tuned using RT-1 training data before the evaluation.
The distinction matters because developers still need to collect suitable robot-specific data and adapt the model before expecting reliable performance on new hardware.
Runway Releases Praxis-1 Under Apache 2.0
The public release is significant because Runway has made both the model weights and the necessary development code available under the Apache 2.0 licence.
The licence permits commercial use, modification and redistribution, subject to its terms.
Developers can access the model through several platforms:
- GitHub: Source code for inference and fine-tuning.
- Hugging Face: Model weights and checkpoints, with native LeRobot support.
- vLLM-Omni: Support for model serving and inference.
- Reactor: Hosted inference options.
The release includes support for full fine-tuning and Low-Rank Adaptation (LoRA). LoRA allows developers to adapt a model by training a smaller set of additional parameters rather than updating the entire model.
This can reduce the resources required for certain customisation tasks.
The open-weight approach also gives robotics developers greater control over where the model runs and how it is adapted to their equipment.
Praxis-1 Supports NVIDIA GPUs and Robotics Hardware
Runway says Praxis-1 can operate across several categories of NVIDIA hardware.
The supported platforms include:
| Hardware | Intended use |
|---|---|
| NVIDIA H100 | Data-centre inference and development |
| NVIDIA B200 | High-performance AI infrastructure |
| NVIDIA RTX 5090 | Workstation-level model execution |
| NVIDIA RTX PRO 6000 | Professional AI development |
| NVIDIA Jetson Thor | On-robot edge computing |
The standard model is relatively demanding to run. Technical documentation indicates that the full BF16 weights require approximately 54.4GB of memory, with an 80GB GPU recommended for standard inference.
Runway has also made a quantised NVFP4 version available that is designed to run on an NVIDIA GeForce RTX 5090.
This gives smaller robotics development teams another potential route to experimentation without relying exclusively on large data-centre systems.
Nevertheless, hardware compatibility does not mean that every robot can immediately use the model without modification. Developers still need to adapt the software to the robot’s cameras, movement system, physical configuration and intended tasks.
Early Robotics Partners Have Already Tested the Model
Before the public release, Runway worked with three robotics partners to test Praxis-1 on different types of hardware.
These included:
- Noble Machines: Testing the model on a legged humanoid robot.
- Standard Bots: Using it with a six-axis industrial robotic arm.
- Ultra: Testing it on a dual-arm warehouse robotics platform.
The partnerships allowed Runway to examine how the same underlying model could be adapted to different robotic configurations.
The company has also demonstrated Praxis-1 performing manipulation tasks, including picking up objects, moving items between containers and handling everyday objects.
These examples are important because robotics models must work beyond controlled benchmark environments. Real-world systems need to handle changing lighting, unexpected object positions, occlusion and physical variations.
Runway says its objective is to develop a general-purpose world action model that can support a wide range of physical applications.
Why Open-Weight Robotics Models Matter
The release comes as technology companies increasingly compete to develop AI systems that can interact with the physical world.
Much of the recent AI race has focused on chatbots, image generation and software-based agents. Robotics introduces another challenge because AI systems must understand physical environments and translate decisions into movements.
Access to open-weight models could help researchers and smaller developers experiment with advanced robot-control technology without depending entirely on proprietary services.
It could also encourage more collaboration between model developers, hardware manufacturers and industrial automation companies.
For manufacturers and warehouse operators, the potential applications include material handling, repetitive assembly, object sorting and other physical tasks.
However, commercial deployment will require more than a promising benchmark score. Reliability, safety, latency, maintenance costs and the ability to recover from unexpected situations will all influence whether these systems become practical.
The Bigger Picture
Runway’s Praxis-1 release highlights the growing convergence between generative video models and physical AI. By using knowledge learned from large-scale video to support robot control, the company is exploring an alternative to training every robotic capability exclusively through physical demonstrations. Making the weights and development tools available under Apache 2.0 could also broaden access to advanced robotics research and commercial experimentation.
However, the technology remains at an early stage of practical deployment. Robot-specific fine-tuning, substantial computing requirements and real-world validation remain important barriers. The long-term significance of Praxis-1 will depend on whether developers can use it reliably across different machines and environments, rather than only reproducing results in controlled evaluations.
Looking Ahead
Runway is making Praxis-1 available to robotics developers and researchers through GitHub and Hugging Face, with additional support for inference and fine-tuning tools. The next stage will involve wider testing across robot platforms, industrial environments and everyday tasks. Developers will be watching whether the model can reduce the amount of specialised training data required to adapt robots to new applications.
The launch also places Runway in the expanding physical AI market, where companies are developing models that connect visual understanding with real-world actions. If video-based pretraining proves effective at scale, it could change how robotics systems are trained and deployed. For now, Praxis-1 represents an important open-weight release, but its commercial impact will depend on practical reliability and adoption beyond early demonstrations.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
