Robotics

Runway Opens Praxis-1: A 27B Video-Trained Robot Model Tops NVIDIA's RoboLab-120 Among World Action Models

The Apache 2.0 model predicts future video and robot actions together, scores 46.2% on RoboLab-120, and runs from data center GPUs down to Jetson Thor on the robot.

DailyMech Staff

Published

Last updated / Fact-checked on

Runway Opens Praxis-1: A 27B Video-Trained Robot Model Tops NVIDIA's RoboLab-120 Among World Action Models
Pixabay — Pixabay Content License

Illustrative robot arms

Illustrative image of robot arms; not a depiction of Runway, Praxis-1, or any specific robot. Image: Pixabay, Pixabay Content License.

Runway, best known for its generative video tools, has released the weights for Praxis-1, a robot-control model built on the company's large-scale video pretraining. The 27-billion-parameter model is available under the Apache 2.0 license, including for commercial use, along with inference and fine-tuning code. Runway first introduced Praxis-1 in September as its first "world action model" and made it broadly available this week.

The release is a bet on a specific idea: that the most scalable source of robotics training data is not expensive teleoperated robot footage, but ordinary video.

What a World Action Model Is

Most open robot foundation models today are vision-language-action models (VLAs), which map camera images and a text instruction to robot motions. Praxis-1 takes a different route. According to Runway's technical report, it is a mixture-of-experts diffusion transformer that predicts what the scene will look like next and the motion that gets it there, at the same time.

The model has three parts. A video backbone predicts upcoming frames. A smaller "action expert" reads the backbone's internal features as it imagines the future and works out the motion. Thin per-robot projections translate each robot's poses and actions into a shared format, so one model can drive different hardware. StartupHub.ai, citing Runway, reports a 26.9B-parameter video backbone paired with a 16-layer action expert.

On each call, Praxis-1 takes an instruction, the latest camera frames and the robot's current pose, and returns a predicted chunk of future video along with 1.6 seconds of actions: 24 action rows at 15 Hz. The robot executes the chunk, sends back what its cameras saw, and calls again over a single WebSocket.

Training on Video First

Runway says training runs in four stages, each building on the last, and video prediction stays part of the objective throughout. It starts from general video, moves through footage of hands and robots, then introduces action targets, and finally fine-tunes for a target robot. According to StartupHub.ai's reading of the report, mid-training used roughly 58% egocentric human video, 20% general video and 18% real robot video. Fine-tuning for the Franka arm used only successful episodes from the DROID dataset, about 277 hours.

Runway claims its research shows "web video matches teleoperated robot video as pretraining data for policy performance," which would shift the bottleneck from collecting robot demonstrations to how much general video a model can learn from.

Benchmark Results

On NVIDIA's RoboLab-120, a simulation benchmark of 120 tabletop tasks for a DROID-style arm run at 10 trials per task, Praxis-1 fine-tuned on DROID solved 555 of 1,200 episodes, a 46.2% success rate. Runway reports that this is the highest of any world action model and any open-source robot foundation model on the leaderboard.

Model RoboLab-120 Success Source
Praxis-1 46.2% Runway
FLUX 3 Action 42.9% Runway
Cosmos3 Nano 36.8% Runway
π0.5 28.0% Runway

There is an important caveat. Times of AI notes that on the leaderboard's overall table Praxis-1 ranks fifth, behind systems including agentic language-model policies at 86.0% and 57.1%. Runway acknowledges those LLM-agent entries in its report. Praxis-1 leads its category, not the whole board.

Runway also reports real-world results. On a physical Franka FR3 arm operated by Positronic, using FLUX 3 Action's published single-item tasks, Praxis-1 succeeded in 88 of 100 episodes, compared with 80 for FLUX 3 Action across the same 10 tasks, according to AI.info. Fine-tuned on Google's RT-1 robot, which it never saw in training, Praxis-1 reached 80.0% on the SimplerEnv test versus 52.4% for the original RT-1 policy, Runway said on X.

Built to Run Anywhere

Praxis-1 is available on GitHub, on Hugging Face with native LeRobot support, and in vLLM-Omni, with hosted inference on Runway's Reactor service. Runway says it runs out of the box on NVIDIA hardware from H100 and B200 data center GPUs to RTX 5090 and RTX PRO 6000 workstation cards, and on Jetson Thor aboard the robot itself. Both full fine-tuning and LoRA are supported, so developers can adapt it to their own embodiment, cameras and tasks without retraining the whole model.

Runway's report also tested a quantized NVFP4 version, which succeeded in 279 of 600 sampled RoboLab-120 episodes (46.5%) versus 287 of 600 (47.8%) for the same checkpoint in bf16, suggesting low-precision deployment costs little accuracy. Launch partners have run it on a humanoid, an industrial arm and a warehouse robot, according to Times of AI.

Why It Matters

Praxis-1 blurs the line between generative video companies and robotics foundation model makers. NVIDIA, Black Forest Labs and Runway are now all pursuing models that jointly predict the visual future and the actions to reach it. An open, commercially licensed model at this scale lowers the barrier for robotics startups and labs that cannot afford to pretrain their own.

Limits to Keep in Mind

RoboLab-120 is a simulation benchmark. NVIDIA says its rankings correlate strongly with a real-world benchmark, but simulation success is not deployment success. Physical AI News also points out that without detail on data volumes and usage terms for the video corpus, the strength of the training foundation is hard to assess independently. Real-world results so far come from tabletop pick-and-place style tasks, not long-horizon factory or home work.

What to Watch

Independent replications on new hardware, community fine-tunes on Hugging Face, and whether the "video is enough" thesis holds for dexterous, long-horizon tasks will determine whether Praxis-1 becomes a standard base model for robotics developers.

How We Reported This

Based on Runway's November 12 announcement and technical report, Runway's post on X, and coverage from Times of AI, AI.info, StartupHub.ai and Physical AI News, reviewed November 13, 2026. All benchmark and performance figures are as reported by Runway. No hands-on testing was conducted.

Published and fact-checked November 13, 2026.

Sources

References

DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.

In Robotics

From other sections

Sections