Robotics News: Runway Opens Praxis-1 Robot Model

The most important robotics news of the week is not a new robot, but a new brain that any robot builder can now download. Runway, the company best known for its video-generation tools, made Praxis-1 generally downloadable on November 11, 2026, releasing the weights and supporting code for its video-trained world-action robot model under the Apache 2.0 licence. The system was first announced in September and tested with a small group of selected partners, so the news is not the model's first appearance but its public opening: from today, researchers, startups and university laboratories can take a powerful robot-control model, adapt it to their own hardware and tasks, and use it commercially without paying a licence fee. In a field where the best robot brains have usually been locked inside corporate laboratories, that is a genuinely significant shift.

Praxis-1 is a roughly 27-billion-parameter model, and its central idea is that a robot can learn a great deal about the physical world simply by watching video. Rather than training a robot only on expensive, hard-to-collect demonstrations of robot movement, Runway trained its system on vast quantities of video, including footage of people doing everyday tasks, and then taught it to turn what it sees into physical actions. The model predicts an upcoming video segment and the actions a robot should take at the same time, combining large-scale video learning with control that is specific to the robot it is running on. The result is a system that can imagine what the world will look like a moment from now, and then choose the movements that will bring that future about.

AI News: How Praxis-1 Learned to Act by Watching Video

The hardest problem in robotics has always been data. A language model can be trained on text scraped from the internet, and an image model can learn from billions of photographs, but a robot model needs examples of physical action, and those examples are slow and expensive to collect. Every demonstration requires a real robot, a real environment and a human operator, and the resulting datasets are tiny compared with the corpora used to train chatbots. Runway's answer is to sidestep the bottleneck: teach the model about the physical world first by showing it video, and only then teach it how to move a robot's joints.

Praxis-1 is built around a 26.9-billion-parameter video backbone, a mixture-of-experts diffusion transformer descended from Runway's video foundation model, paired with a much smaller action expert, a 16-layer transformer of roughly 300 million parameters that reads the video backbone's internal state as it imagines the future. Together the system totals about 27.2 billion parameters, although only around 2.8 billion parameters are active for each token the model processes, which keeps the computational cost of running it far below what the headline number might suggest. On each control call, the model takes in a language instruction, recent camera frames and the robot's current joint positions, and jointly predicts a future video chunk and 1.6 seconds of motor commands, expressed as 24 action steps at 15 hertz.

The training recipe is where the video-first philosophy shows most clearly. According to Runway's technical report, the training mix was 58 percent egocentric human video, meaning footage shot from a person's point of view as they perform tasks with their hands, 20 percent general video and 18 percent real robot video. That blend was then refined through several training stages, including a joint mid-training phase in which the action expert was attached to the video backbone and the two were trained together with equal weight given to video prediction and action prediction. The model was mid-trained across 18 different robot embodiments, and it was fine-tuned on roughly 277 hours of successful episodes from the DROID dataset for the Franka robot arm. The result is a system that has absorbed a broad understanding of how objects move, how hands grasp and how tasks unfold, and can then specialise that understanding for a particular robot.

Robotics News: The Benchmarks and What They Really Mean

Every robotics news story this week comes with a benchmark number, and for Praxis-1 the headline figure is 46.2 percent. Runway reports that the DROID fine-tuned version of the model completed 555 of 1,200 episodes on NVIDIA's RoboLab-120 benchmark, a simulated suite of 120 tabletop manipulation tasks with 10 trials each, for a success rate of 46.2 percent. That puts it ahead of the cited FLUX 3 Action baseline at 42.9 percent, NVIDIA's own Cosmos3 Nano at 36.8 percent and Physical Intelligence's pi 0.5 at 28.0 percent. Among world-action models and open-weight robot policies, Runway says, Praxis-1 now leads the field.

Two caveats belong next to those numbers. First, every figure so far is the company's own: no independent laboratory has yet reproduced the leaderboard result, and the comparison is against other vendors' published figures rather than a head-to-head test run by a neutral third party. Second, a first-place finish at 46.2 percent still means failing more than half the time, and on simulated tabletop tasks rather than a factory floor or a kitchen. The benchmark also shows how far the field still has to go: the same leaderboard lists zero-shot frontier vision-language harnesses at far higher scores, with AMP1 at 86.0 percent and SenseNova-RoboRSI at 57.1 percent, although those systems work very differently from an open-weight model that a laboratory can download and run locally. The honest reading of the RoboLab-120 result is that Praxis-1 is the strongest openly available world-action model yet, and that the strongest openly available model still fails most of the tasks it is given.

Technology News: What Is Actually in the Download

The release is unusually complete for an open model, and the contents of the download may matter more than any single leaderboard position. Three checkpoints are available on Hugging Face: the base mid-trained model, a version fine-tuned on the DROID robot dataset and one fine-tuned for the bimanual YAM robot, which uses two arms working together. Alongside the weights come inference code and both full and LoRA fine-tuning code, published on GitHub under the repository name runwayml/praxis-1-oss, with native support for the LeRobot robotics library and integration paths including serving through vLLM-Omni and hosted inference on Runway's Reactor platform. The Apache 2.0 licence explicitly permits commercial deployment without royalties or usage restrictions.

The practical question for most builders is whether they can run it at all, and the answer is demanding but not out of reach. The full model is 54.4 gigabytes in bfloat16 precision and needs a single 80-gigabyte graphics card to run, which puts it within reach of a well-equipped laboratory or a cloud instance but far beyond a laptop. A 4-bit NVFP4 quantised version shrinks the weights to 17.5 gigabytes, making deployment considerably more practical. Runway says the model runs out of the box across NVIDIA hardware, from H100 and B200 data-centre cards to RTX 5090 and RTX PRO 6000 workstation cards, and on the Jetson Thor computer that sits inside some robots themselves. Separate inference and runtime costs remain, and developers must still bind camera observations and joint state to the model's expected format, select an action head for their robot's embodiment and often fine-tune on local demonstrations before the system will do anything useful.

Artificial Intelligence News: Why an Open Robot Brain Matters

The significance of Praxis-1 lies less in any single capability than in what its openness makes possible. Robot learning has long been held back by the scarcity of training data and by the concentration of the best models inside a handful of large companies. An open-weight release under a permissive licence changes the economics for everyone else. Small laboratories, university groups and startups can now fine-tune a strong starting model for their own hardware and tasks instead of starting from zero, and they can inspect, adapt and serve a shared policy rather than relying on a black-box endpoint whose behaviour they cannot examine. That could expand experimentation across industrial arms, low-cost research robots and warehouses, while also distributing responsibility for physical safety to the teams that deploy the model, which is a serious obligation that comes with the freedom.

The release also fits into a broader pattern in this week's artificial intelligence news, in which open-weight models are drawing fresh attention alongside the closed systems that dominate much of the OpenAI news cycle. Runway's bet is that openness will accelerate progress faster than secrecy, and that the community's ability to test, break and improve a shared model will outweigh the risks of putting powerful robot-control weights into the public domain. That bet is not without critics: a model that can control a physical machine raises safety questions that a text generator does not, and the same weights that let a warehouse robot pick up a box could, in the wrong hands, be adapted for harmful purposes. The Apache 2.0 licence contains no restriction on such use, which is why Runway's emphasis on guarded testing and staged validation matters as much as the licence itself.

What Happens Next for Praxis-1

The immediate next step belongs to the community. Researchers and builders will download the weights, try them on their own robots and report what works and what does not, and the first independent evaluations will be the real test of Runway's claims. Watch for results on real hardware rather than in simulation, because a policy that scores well on RoboLab-120 may behave very differently when it has to cope with imperfect lighting, worn grippers and the unpredictability of a real room. Watch also for fine-tuned variants: the LoRA fine-tuning code is designed to let small teams adapt the model cheaply, and the most interesting early results may come not from the base model but from versions specialised for particular robots and tasks, from low-cost arms used in university teaching to industrial manipulators in warehouses.

The longer-term question is whether video pretraining becomes the standard recipe for robot learning. If it does, the bottleneck in robotics shifts from collecting robot demonstrations to curating video, and the companies with the largest video libraries gain a lasting advantage. Runway, whose business began with video generation, is well placed in that world, and its decision to open the weights suggests it believes the ecosystem around Praxis-1 will be worth more than the model alone. Competitors will now have to decide whether to match that openness or to compete on performance behind closed doors, and the answer will shape the robotics industry for years.

Source: SXF Brief