Runway Unveils ‘Praxis-1,’ Linking Video Pretraining to Robot Control

Runway unveiled Praxis-1, a world action model that links video pretraining to robot control, on November 11, 2026. The 27B-parameter model’s weights, along with code for inference and fine-tuning, are available under the Apache 2.0 license, which permits commercial use. Runway said the model achieved a 46.2% success rate in an evaluation of 1,200 episodes on NVIDIA’s RoboLab-120.

Runway released Praxis-1’s weights and inference and fine-tuning code under the commercially permissive Apache 2.0 license, and says the model runs on multiple generations of NVIDIA products. For robot developers, it offers an option to adapt a video-pretrained model to their hardware and data instead of building one from scratch. However, benchmark results alone do not establish safety or performance in real-world operation.

What stands out about Praxis-1 is that it not only repurposes a video-based foundation model for robot control, but also predicts video and actions together in one model and releases its weights and fine-tuning code. Although the Apache 2.0 license permits commercial use, the release itself does not mean that the model has been validated for operation or safety on different robots. The 46.2% result on RoboLab-120 is a benchmark score across 1,200 episodes and should be distinguished from success rates in the field. On the data front, Runway points to research findings that web videos can serve as pretraining data on par with videos of teleoperated robots, and identifies the volume of general video a model can learn from as the bottleneck for capable policies. This is a different layer of work from Memories.ai reportedly building a standardization pipeline for first-person video data for robots, which also addresses the collection and organization of video data. A video-collection pipeline and a model that learns actions from that video are connected, but without information on data volume and usage terms, the effectiveness of the training foundation cannot be assessed. The model release and claimed support for hardware ranging from data-center GPUs such as H100 and B200 to Jetson Thor point to the possibility of using the same model family from training and fine-tuning through inference on robots. However, hardware support claims alone do not reveal processing speed, latency, memory requirements or operating conditions on actual robots. As with UC Berkeley and NVIDIA reportedly releasing an AI agent development tool for robotics, the release of a model or tool should be evaluated separately from its reproducibility on real-world hardware. The next points to verify are how much adaptation is needed for different robot camera configurations and tasks, performance in unseen environments, and safety and sustained operation on real robots. The 80.0% result in SimplerEnv using RT-1 is one example of adaptability, but Runway’s public information does not specify real-robot evaluation conditions, operating duration or behavior in the event of failure. The key question is how the company’s claim that open weights give developers the freedom to inspect and modify the model translates into reproducibility across hardware and validation in real-world use.

Praxis-1 is a mixture-of-experts diffusion transformer that predicts video and actions simultaneously. Runway said FLUX 3 Action scored 42.9%, Cosmos3 Nano 36.8% and π0.5 28.0% on RoboLab-120, and described Praxis-1 as the top performer in the evaluation. The model is available through GitHub, Hugging Face and vLLM-Omni, and supports full-parameter fine-tuning and LoRA. Runway said it runs on NVIDIA H100, B200, RTX 5090, RTX PRO 6000 and Jetson Thor. The company also said that when adapted to robots not included in Google RT-1’s training data, Praxis-1 achieved an 80.0% success rate in SimplerEnv, exceeding the original RT-1 policy’s 52.4%.