On November 11, 2026, Runway, the company known for its video generation AI, released the weights and code for Praxis-1, an AI model that generates robot actions from instructions and camera images. It is available under the Apache 2.0 license, which permits commercial use, so developers can run the model freely and fine-tune it for their own purposes. Partner companies had been testing the model since its announcement in September, and it is now open to outside developers as well.
Praxis-1's defining feature is that it applies the ability of video-prediction AI, which forecasts what happens next in a video, to robot control. Because knowledge learned from general video can be put to use in generating robot behavior, developers don't have to build a model from scratch. However, the model still has to be fine-tuned for each target robot, and the success rates Runway published are results from simulation evaluations of models that had already been adjusted this way.
Model weights and code released under Apache 2.0
Runway announced Praxis-1 in September, and three companies, Noble Machines, Standard Bots and Ultra, had been testing it. With the public release, the code needed for inference and fine-tuning is available alongside the model weights. Two fine-tuning methods are supported: full fine-tuning, which updates all of the model's parameters, and LoRA, which adds a small number of parameters and adjusts only those.
Runway lists GitHub and Hugging Face as distribution channels, along with LeRobot, vLLM-Omni and Reactor. Developers can use the pretrained model as a foundation and tailor it to their own robots and tasks. Still, being able to run the model is not the same as having a robot that can perform a target task. Practical use requires adapting the model to the specific robot and evaluating its performance in the actual work environment.
How video prediction is applied to robot control
Praxis-1 is a "world action model" that predicts future video and robot actions at the same time. At its core is a foundation model responsible for video prediction. An action-generation module of about 314 million parameters reads its internal representations and converts them into robot movements. The model also includes components that translate the differing inputs and outputs of each robot, bringing the total scale to about 27 billion parameters.
In its published technical report, Runway explains that a small action-generation module draws on the internal state of a video model that predicts future frames. The model takes a task instruction, camera images and the robot's pose as input. By default, it generates 24 steps of motion at 15 Hz, equivalent to 1.6 seconds, and outputs 21 steps on the first inference.
Training proceeds in several stages. The model is first pretrained on general video, then trained further on video showing human hands and robots. Video prediction and action generation are then trained jointly, and finally the model is fine-tuned on data from the target robot. By preserving video-prediction capability even while learning action generation, the design aims to carry knowledge gained from video over to robot control.
That said, a robot cannot learn to operate simply by watching video. For example, fine-tuning the DROID model used about 277 hours of demonstration data of successful manipulation. Training used 128 NVIDIA H100 GPUs and 50,000 update steps. Deploying the model on a new robot requires collecting demonstration data suited to that machine and securing the computing resources for fine-tuning.
The 46.2% and 80.0% success rates come from different evaluation conditions
In the performance evaluations Runway published, the DROID model recorded a 46.2% success rate on RoboLab-120, and the RT-1 model recorded 80.0% on SimplerEnv. Both figures, however, are results after fine-tuning on data from the target robot, and the simulation environments and tasks used for evaluation also differ. The conditions described in Sections 6 and 8.1 of the technical report are summarized below.
| Model / evaluation | Fine-tuning conditions | Environment / number of trials | Success rate (reported by Runway) |
|---|---|---|---|
| DROID model / RoboLab-120 | About 277 hours of successful demonstrations, 128 H100 GPUs, 50,000 updates | Simulation, 1,200 trials | 46.2% |
| RT-1 model / SimplerEnv | All RT-1 training data (about 78,000 episodes, 95 hours); trained for about 28 hours and 30,000 steps on 32 H100 GPUs | Simulation under the visual matching condition, 4 tasks, 864 trials | 80.0% |
These are simulation results reported by Runway itself and do not show that the same success rates would be achieved on real robots. Because the robots, tasks and trial conditions differ, the 46.2% and 80.0% figures also cannot be compared directly to judge which model is better.
The evaluation conditions for the RT-1 model deserve particular attention. RT-1 was not among the robots included in Praxis-1's pretraining or mid-stage training, but the model was fine-tuned on RT-1's training data before evaluation. In other words, the 80.0% success rate was not obtained by applying an unadjusted model directly to an unseen robot. Runway also reports that the original RT-1 model achieved a 52.4% success rate in the same SimplerEnv evaluation.
These results suggest that Praxis-1, pretrained on video, may be applicable to controlling different robots. At the same time, they make clear how much fine-tuning is needed in practice. The released model is a starting point for development, but achieving comparable performance on a given robot and work environment will require individual adjustment and validation.
Inference requires an 80GB GPU; a lighter model is also in development
Hardware requirements matter for anyone using Praxis-1. According to the technical report, the standard bf16 model weights are 54.4GB and run on a single GPU with 80GB of memory. Typical inference on an H100 takes about 0.55 to 0.6 seconds and produces motion equivalent to 1.6 seconds. By default, the model generates multiple actions at once and executes them in sequence. These are inference-time conditions and should be distinguished from the computing resources needed for fine-tuning described above.
To reduce computational load, Runway quantized the model's routed experts, specialized modules used selectively depending on the input, to NVFP4 format and evaluated a fine-tuned model that way. Its weights total 17.5GB, considerably smaller than the standard bf16 format. It succeeded in 279 of 600 trials, a 46.5% success rate. Running the same fine-tuned checkpoint in bf16 format produced 287 successes, or 47.8%. The technical report says an NVFP4 checkpoint will be released in the future, but that statement alone does not confirm whether it is currently available.
Runway has also published examples of the model operating real robots. The technical report shows the dual-arm robot YAM and the single-arm robot Franka placing a cup upside down and closing the lid of a box, in footage played at 1x speed. In each case the work was done without teleoperation. However, these are demonstrations of real-world operation and, unlike the simulation evaluations above, are not success rates calculated from many trials.
Challenges remain. Runway says the model can mix up objects that look alike and may fail to interpret ambiguous instructions correctly. For tasks made up of multiple steps, it can carry them out when given a separate instruction for each step, but completing a whole sequence from a single instruction is difficult. Long-horizon planning and self-improvement are positioned as future research topics.
With Praxis-1 now public, developers can build robot control systems on top of a large model trained on video. But what has been shown so far is simulation performance achieved through robot-specific fine-tuning, plus limited real-world demonstrations. Going forward, two things will be key indicators of practicality: how far the data and computing resources needed for fine-tuning can be reduced, and how reliably the model can carry out multi-step tasks in real environments.