Praxis-1 builds on Runway’s work as a founding member of the Cosmos Coalition, advancing open world models for physical AI. | Source: Runway
Runway AI Inc. yesterday released Praxis-1, its world action model for robotics. The model is now available to developers with open weights under the Apache 2.0 license.
Runway first announced Praxis-1 in September with launch partners including Noble Machines, Standard Bots, and Ultra. Since then, these companies have been running the model on their own hardware, including a legged humanoid, a six-axis industrial arm, and a dual-arm warehouse robot.
“A strong base model goes a long way. Earlier, with Noble Machines, our previous model needed a lot of training to learn their warehouse task: picking up wooden planks one at a time and placing them in a container,” Andy Chen, the head of robotics at Runway, told The Robot Report. “Praxis-1, built with more compute and data, learned it from half an hour of examples and got it right 15 times out of 15. Then we gave it a messy pile of planks it had never seen to clear into bins, and it cleared the whole pile six times out of 10.”
“Iterating on feedback from partners makes our models better,” he continued. “Today, Praxis-1 is the top-ranked world action model on NVIDIA’s RoboLab leaderboard, and the top-ranked open-source, end-to-end policy.”
Praxis-1 is available on GitHub, on Hugging Face with native LeRobot support, and in vLLM-Omni, with hosted inference on Reactor. Runway said it released everything developers need to run and adapt the model on their own hardware under Apache 2.0, including for commercial use. That covers the weights and the inference and fine-tuning code, with support for both full fine-tuning and low-rank adaptation (LoRA).
Praxis-1 runs out of the box across NVIDIA hardware, from data center GPUs such as the H100 and B200, to workstation GPUs like the RTX 5090 and RTX PRO 6000, to Jetson Thor on the robot itself.
Praxis-1 builds on Runway’s video foundation models
Praxis-1 is a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together.
On RoboLab-120, Praxis-1 reached 46.2% success across 1,200 episodes, compared with 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5. It ranks fifth overall on the leaderboard.
“The core piece is that a model trained to predict video really understands a lot about the physical world: how objects behave, how hands move, what a task looks like partway through, and so on,” Chen said.
“Praxis-1 is built on Runway’s video foundation model, and it learns how the world works before it learns to act,” he added. “We keep video prediction in the training objective the whole way through, so the model pictures what the scene should look like next and produces the motion that gets it there.”
Praxis-1 works across different robot embodiments
Praxis-1 learns how objects behave and how tasks unfold from large-scale video. | Source: Runway
Praxis-1 learns how objects behave and how tasks unfold from large-scale video, which gives it a head start on a new embodiment before fine-tuning on any hardware or domain-specific data, explained Runway. The company’s research has shown that web video matches teleoperated robot video as pre-training data for policy performance, which shifts the bottleneck on a capable policy to how much general video a model can learn from.
“The tasks vary with the robot. Noble Machines builds a legged humanoid for heavy industrial handling, and with them Praxis-1 picks up wooden planks, places them in containers and clears mixed piles into bins,” Chen said. “Standard Bots makes six-axis arms for factory work like machine tending and palletizing. Ultra builds and deploys dual-arm stations that pack orders in warehouses. A humanoid, a single arm, and a pair of arms are very different machines, and the same model drives all three.”
“With our partners, a single Praxis-1 checkpoint has adapted to a variety of embodiments with light fine-tuning, and with LoRA, you can adapt it to your own embodiment, cameras and tasks without retraining the full model,” said Runway. “Fine-tuned on Google’s RT-1 robot, which never appeared in its training data, Praxis-1 reaches 80.0% success on SimplerEnv, well above the original RT-1 policy at 52.4%.”
Submit your innovation now!Why keep Praxis-1 open weight?
Runway has been committed to shipping Praxis-1 as an open weight model since its initial announcement. Open weights give hardware developers the flexibility to run, inspect, and adapt a world action model on their own terms, according to the New York-based company.
“At Runway, we believe that U.S. leadership in physical AI is critical,” Chen asserted. “Regaining our global lead in manufacturing, accelerating productivity across our economy, and securing ourselves and our allies will rely on this.”
“Achieving this will require a significant increase in investment across both hardware and software, but it will also require a level of interoperability and openness from American models that does not currently exist for physical use cases,” he said. “We view open world models as a compounding advantage that gives hardware developers flexibility and control they don’t currently have.”
After making Praxis-1 available to developers, Runway said it is focusing on scaling deployments.
“We want Praxis-1 in the real world: in real warehouses, factories, and industrial sites, doing real work,” Chen said. “We believe the frontier of robotics will be discovered through real-world applications, not in a lab. So what comes next is deploying with our partners, learning from what the model runs into out there, and putting those lessons back into the model.”
Tell Us What You Think!