Newshot
Account

Runway's Praxis-1 robot model beats GPT-6 Astra-powered arm: 15.6x faster and 1/51 the cost

AI summary

US video AI company Runway announced Praxis-1, an open-weight robot model under the Apache 2.0 license that applies knowledge learned from video pretraining to real robot control, letting developers run and fine-tune it on their own hardware. Built on Runway's 27-billion-parameter video foundation model, it predicts both the next scene state and the actions needed to achieve it, and is available in an NVFP4 quantized version that runs on Nvidia RTX 5090 GPUs, with weights and code published on Hugging Face and GitHub. On the RoboLab-120 benchmark it completed 555 of 1,200 scenarios (about 46.2% success), and Runway says it completes tasks 15.6 times faster than an arm driven by GPT-6 Astra at 1/51 the cost per task.

Translation(model translation, not the source's own text)

The strongest AI engine for robotic arm tasks: Praxis-1 beats GPT-6 Astra outright, 15.6 times faster and at 1/51 the cost

IT Home, November 12: US video AI company Runway announced yesterday (November 11) the launch of its robotics model Praxis-1, which applies knowledge learned from video pretraining to real robot control. The model has open weights and is released under the Apache 2.0 license, so developers can run and fine-tune it on their own hardware.

The design approach behind Praxis-1 is to have the AI first learn from ordinary video how objects move and how human hands manipulate things, and then apply that knowledge to robot motion learning, thereby addressing the high cost of collecting robot training data. When executing a task, the model not only predicts what the scene will look like next, but also predicts which actions are needed to achieve that outcome.

The model is built on Runway's 27-billion-parameter video foundation model, supports local operation, and offers an NVFP4 quantized version that can run on NVIDIA RTX 5090 graphics cards. The model weights and related code are published on Hugging Face and GitHub respectively.

According to test results disclosed in reports, Praxis-1 completed 555 of the 1,200 scenarios in the RoboLab-120 benchmark, for a success rate of about 46.2%, leading in the world action model and open robot foundation model categories. Runway said the research is still at an early stage, and future work will focus on long-horizon planning and on enabling the model to use its predictions to evaluate and improve its own actions.

The company says Praxis-1 is significantly better than the current state-of-the-art laser robotic arm in both speed and cost. On the same workstation and tasks, Praxis-1 completes tasks 15.6 times faster than a robotic arm driven by GPT-6 Astra, at 1/51 the cost per task.

Advertising statement: External redirect links contained in this article (including but not limited to hyperlinks, QR codes, passcodes and other forms) are used to convey more information and save selection time; the results are for reference only. All articles on IT Home that contain external links include this statement.

Related articles

  • China's first fully automatic verification device for embodied intelligence pressure gauges has been successfully developed: efficiency improved by more than 3.5 times, led by Sinopec
  • NIO's Li Bin: the most shipped embodied intelligence agent today is the smart electric vehicle
  • JD.com discloses progress in embodied intelligence data collection: covering more than 100 segmented scenarios and collecting over 100,000 real work tasks
  • UBTECH's Thinker-Sim embodied intelligence simulation platform released, providing a one-stop system for robot capability evaluation
  • Embodied intelligence company Wujing Dynamics reportedly halts work and owes nearly 8 million yuan in wages: CEO says the actual controller cannot be reached; the company won nearly 1 billion yuan in orders just 4 months after founding
  • "Gamma Intelligence," an embodied company incubated by China Telecom, is established to advance the "self-evolution" of humanoid robots

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗