FRIDAY, NOVEMBER 20, 2026 Tech news, minus the noise.
N News Cast
LATEST
Home  /  AI
AI

Runway's Praxis-1 Teaches Robots by Watching the World, and the Weights Are Free

Runway opened up Praxis-1, a 27-billion-parameter robot model that predicts video and robot actions together, under an Apache 2.0 license.

Runway's Praxis-1 Teaches Robots by Watching the World, and the Weights Are Free
Source: Jo Teichmann / KUKA (public domain), via Wikimedia Commons

Runway Opens a Robot Brain Anyone Can Download

Runway made Praxis-1 generally available on Wednesday, November 11, and the release is more generous than most robot model launches. Praxis-1 is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together, and it ships with open weights under the Apache 2.0 license, including for commercial use. That means a robotics team can download it, fine-tune it, and put it in a product without asking Runway for permission or paying a license fee.

The idea behind the model is simple to state and hard to build. Most robot models learn from teleoperation, where a human drives a robot arm through a task again and again while the system records every joint movement. That works, but it is slow and expensive, and every new task needs new demonstrations. Praxis-1 tries a different route: it learns from video of the world, then connects what it sees to the actions a robot should take. Watch enough footage of hands opening drawers and doors, and the model starts to understand the shape of those jobs before it ever touches a handle.

What the Praxis-1 Robot Model Actually Does

Praxis-1 is a world action model, which is a mouthful for a system that does two things at once. It predicts what the next frames of a video should look like, and it predicts the robot actions that would produce them. Tying those two jobs together is the point. A model that only predicts actions can learn shortcuts that look right in training and fail on a real robot. A model that also has to predict the video has to understand physics, contact, and cause and effect, because the frames it generates have to stay plausible.

Runway reports a score of 46.2% on RoboLab-120, a benchmark for robot manipulation. That number needs context. Robot benchmarks are still young, and a single lab score says little about how a model behaves in a messy kitchen or a warehouse aisle. But the direction matters: an open model scoring competitively on a public benchmark gives every robotics lab a starting point that did not exist a month ago. Before Praxis-1, the serious robot models mostly lived behind company APIs or partnership programs.

Why Open Weights Change the Robot Race

The robot field has been splitting into two camps. On one side are the closed systems: Google DeepMind released Gemini Robotics 2.5 to robotics partners on November 10, and NVIDIA keeps pushing its GR00T foundation models through its own ecosystem. These are powerful, but access depends on being a chosen partner. On the other side is a growing open stack, and Praxis-1 is now the most capable open entry in it.

Open weights matter more in robotics than in chatbots. A language model runs the same everywhere, but a robot model has to meet a specific arm, a specific gripper, and a specific workspace. Teams need to fine-tune, and fine-tuning is only practical when you hold the weights. Apache 2.0 licensing means startups can build on Praxis-1 without worrying that the license will change under them, which is exactly the kind of certainty a hardware company needs before betting a product line on someone else's model.

There is a catch, and it is the same catch that follows every open model release. The weights are free, but the compute, the robots, and the safety testing are not. Running a 27-billion-parameter model on a robot in real time takes serious onboard hardware or a fast link to a server. And a robot that acts in the physical world can break things and hurt people in ways a chatbot cannot. Open access does not remove the need for careful testing.

The Honest Limits of Praxis-1

Praxis-1 is a research-grade release dressed as a product launch, and buyers should read it that way. A 46.2% benchmark score means the model fails more often than it succeeds on the test's hardest tasks. Video prediction is a useful training signal, but watching a task is not the same as doing it, and the gap between the two is where most robot demos die. The model also inherits the usual problem of learned systems: it is only as good as the data it trained on, and unusual objects, bad lighting, and cluttered scenes will still trip it up.

What makes this release worth watching is not a claim that robots are solved. It is that the tools to work on the problem are now in more hands. A university lab, a warehouse automation startup, or a hobbyist with a robot arm can all start from the same checkpoint this week. That is how fields move fast, and it is why an open robot model is bigger news than another point on a benchmark chart.

For context on where robot models stood before this release, see our coverage of NVIDIA's GR00T 2, which teaches robots by watching human video, and the Edge0-145B open weights release, which pushed large open models onto a single GPU. The open model wave that started with K2-Horizon now has a robot in it.

AS
News Cast Editorial Team

We cover AI, gadgets, startups and security — original reporting and explainers, written by humans, for humans. Got a tip? Reach us via the contact page.