Runway leaps into robotics with Praxis-1 and launches an ad agent

Praxis-1 is an open-weight world action model from Runway designed to help developers control physical robots with just a bit of light fine-tuning.

Runway co-CEO Cristóbal Valenzuela unveiled the model on Wednesday during the Runway AI Summit. Alongside it, the company launched Runway Ads, an end-to-end agent that can generate, edit, publish, and track paid advertising campaigns completely autonomously.

What it does and how it works

Praxis-1 applies Runway's years of video generation expertise directly to physical robotics.

According to Valenzuela, scaling laws apply to video just like language models. While an LLM predicts the next token, a video model predicts the next frame. By learning to simulate how the physical world behaves on screen, large-scale video models act as generalization engines that can be adapted into control policies for physical hardware.

On the commercial side, Runway Ads marks the company's move into autonomous enterprise agents. Instead of just delivering raw clips, the agent handles the full media buying lifecycle from asset creation to performance measurement.

The reality check on physical AI

Despite dropping the open weights, fully live production deployments of physical AI remain 12 to 18 months away.

The real hurdle is not training data, but runtime speed and inference optimization. Robots need real-time, ultra-fast responses in the real world, which requires significant engineering before heavy world models can power physical machines reliably. Valenzuela noted that Praxis-1 is merely the first generation.

Valenzuela also warned that the US might be "LLM-pilled" by focusing disproportionately on coding and text models. In contrast, he estimated that China directs around 70% of its token consumption toward video models and is moving significantly faster in robotics, creating an urgency for Western teams to build strong foundational world models.

Why it matters

For robotics engineers and researchers, Praxis-1 offers an open-weight foundation that eliminates the need to build a physical simulator from scratch. It gives developers a practical starting point to test whether video-derived spatial understanding can reliably transfer to mechanical arms or quadrupeds.

We think Runway is making a clever strategic pivot by turning its media assets into real-world control systems. However, bridging visual imagination with mechanical precision is notoriously hard. Whether next-frame prediction can truly compete with traditional reinforcement learning in mission-critical robotics will face a tough test over the next 12 to 18 months.