發生了什麼事
Runway AI has introduced Praxis-1, a new 'world action model' designed to translate video into actionable control policies for robotic systems. Unlike traditional robotics models that rely heavily on scarce, expensive, and manually collected robot demonstration data, Praxis-1 is built upon the same large-scale video pretraining architecture used in Runway’s existing world models. By training on vast amounts of third-person video, the model aims to understand object behavior and task progression, allowing it to generalize across different robotic embodiments, including bimanual arms and humanoid robots, with minimal .
Runway AI announced Praxis-1, a model that bridges the gap between generative video and physical robot control. The model is designed to function as a generalist policy, meaning it is intended to work across various hardware configurations without requiring extensive retraining for each specific robot.
The model is currently in an early-access phase with select partners. Runway CTO Kamil Sindi noted that the model has demonstrated success in tasks ranging from simple pick-and-place operations to more complex manipulations involving deformable objects, such as packing gift bags.
Runway claims that its simulation-to-real-world correlation is high, at 0.95, which the company asserts is more cost-effective and accurate than traditional 3D reconstruction-based training methods. The company plans to release the model with open weights to encourage broader adoption among hardware developers.
為什麼這很重要
The primary bottleneck in developing generalist robotics is the lack of high-quality, diverse physical interaction data. By shifting the training paradigm from robot-specific demonstrations to general video data, Runway aims to bypass the data scarcity issue that currently limits the scalability of physical AI. According to the company, its simulation techniques for these policies show a 0.95 correlation with real-world results, suggesting a more efficient path to deploying capable robots. Furthermore, Runway’s commitment to releasing Praxis-1 with open weights represents a strategic move to provide hardware developers with greater control and flexibility, which the company argues is essential for advancing domestic manufacturing capabilities in physical AI.
The robotics industry has long struggled with the 'data wall'—the difficulty of collecting enough high-quality, diverse physical data to train truly general-purpose robots. By utilizing the massive, readily available corpus of human video data, Runway is attempting to solve this by teaching models the physics of the world before they ever interact with a physical machine.
The decision to provide open weights is a significant industry move, contrasting with the trend of closed-source models in the space. Runway frames this as a matter of national interest, arguing that open-weight models are necessary for U.S. hardware developers to maintain competitiveness in manufacturing and physical AI.
The ability of a single model to adapt to different embodiments—from bimanual arms to humanoids—with only 'light ' could significantly reduce the time and cost required to deploy autonomous robots in varied industrial and commercial settings.
互動機制：它實際上是如何運作的
以互動方式探索這項發展背後的基礎技術。
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
接下來看什麼
Runway is currently testing Praxis-1 with early partners, including Noble Machines, Standard Bots, and Ultra, who are evaluating the model on their respective hardware. The company has not yet announced a specific date for general availability, stating only that a public release is planned for the coming months. Future updates will focus on expanding the partner program to further evaluate the model's efficacy and safety across a wider range of environments and embodiments. Interested parties should monitor for the official public release of the model weights and any subsequent performance benchmarks released by the company or its partners.
The transition from the current partner-testing phase to a public, open-weight release is the most critical milestone. Runway has not provided a specific timeline beyond 'the coming months.'
The company intends to continue evaluating the model's safety and efficacy across diverse environments. Observers should look for independent verification of the model's performance on hardware outside of the initial partner group.
As Runway expands its partner program, the specific hardware requirements and the extent of the 'light ' required for new embodiments will be key indicators of the model's true versatility and ease of integration for third-party developers.