Runway releases robotics model Praxis-1, saying internet-video pretraining approaches specialized data in object-placement tests
BIBIBI
AT A GLANCE
Runway says that Praxis-1 pretraining on internet video achieved error in object-placement tests close to that from training on robot teleoperation video.
Article
Runway trains robots on internet video, achieving results nearly on par with specialized data
Beating AI Newsflash: Runway has released its robotics model Praxis-1, applying its ability to train AI video models to real robots. It first learns how objects move and how people complete tasks from large volumes of ordinary video, then turns this experience into robot actions.
Runway conducted an object-placement test. After pretraining on internet video, the robot's mean error was 16.1 cm; using specially collected robot teleoperation video instead, the error was 16.0 cm. The test comprised 93 evaluation sets, and the results from the training approaches were nearly tied.
If this approach can continue to scale, the data most lacking for robot training could suddenly increase by several orders of magnitude. Collecting data through human teleoperation of robots requires equipment, facilities, and labor; ordinary video is already widespread across the internet. Runway aims to turn video that originally could not directly control robots into training material for robots to learn actions.Praxis-1 has already been tested on a range of robots from Noble Machines, Standard Bots, and Ultra, including robot arms, dual-arm robots, and mobile robots. Runway also demonstrated the same policy continuing to perform tasks after moving from a studio to a kitchen. The model is currently available only through applications for early testing; its weights are planned for release in the coming months, while its parameter count and more comprehensive independent evaluations have not yet been disclosed.
The original article states that Praxis-1 first learns object motion and how humans complete tasks from ordinary video, then translates this experience into robot actions.
03
Across 93 object-placement evaluations, mean error with internet-video pretraining was 16.1 cm, while error when trained on specially collected robot teleoperation video was 16.0 cm.
04
Praxis-1 It has been tested on a range of robots from Noble Machines, Standard Bots, and Ultra, including robot arms, dual-arm robots, and mobile robots.
05
Runway demonstrated the same policy continuing to perform tasks after moving from a studio to a kitchen.
06
The model is currently available only through applications for early testing, and its weights are planned for release in the coming months.
07
Its parameter count and more comprehensive independent evaluations have not yet been disclosed.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Traditional robot training often uses data generated by humans teleoperating robots.Praxis-1 attempts to first use ordinary internet video to learn patterns of action, then apply the learned experience to control real robots. In the single test reported in the original article, the mean errors associated with the types of training data differed by only 0.1 cm, but this alone does not prove that they perform equally across all tasks and environments.
Why it matters to readers
If internet video can be used reliably for robot training, the vast amount of ordinary video that cannot directly control robots could become training material, reducing reliance on equipment, facilities, and labor for collecting teleoperation data. The original article has not yet demonstrated that this effect can be broadly reproduced.
For newcomers, the key point is that robot training may gain a more accessible source of data; what is currently available are Runway's limited disclosed test results, which are insufficient to determine whether the model is mature or has fully caught up with professional data.
Risks and unknowns
16.1 centimeters and 16.0 centimeters were independently replicated by a third party is not stated in the original text.
The model's parameter count and more comprehensive independent evaluations have not yet been disclosed.
The original text does not provide the complete 93 sets of evaluation task setups, statistical scope, or detailed results.
Related Developments
Loading event timeline…
It remains unconfirmed whether the internet-video training method can maintain its performance across more tasks, robots, and real-world environments.
The exact date when the weights will be released has not been announced.
Related concepts
Pretraining
This term is not in the glossary yet. Browse related concepts in the glossary.
