Runway's Praxis-1: Can Video AI Actually Control Real Robots?

From pixels to physical movement
Runway, primarily known for generative video, is pivoting its research into physical automation with the release of Praxis-1. Unlike traditional robotics models that rely heavily on massive, expensive, and manually curated robot datasets, Praxis-1 leverages the spatial and physical understanding gained from training on general video footage.
By processing camera inputs alongside text instructions, the model calculates the necessary robot movements to manipulate objects. According to The Decoder, the model currently holds a 46.2% success rate on Nvidia’s RoboLab-120 benchmark, positioning it as a leader in the 'World Action Model' category.
How to implement Praxis-1
The model is available under an Apache 2.0 license, allowing for commercial integration. To get started, you will need significant local compute resources.
- Access: Download the weights and source code via the official Runway GitHub repository and Hugging Face.
- Hardware Requirements: The base model requires a GPU with at least 80 GB of VRAM. A compressed version may run on hardware like an RTX 5090 with 32 GB of VRAM, though performance benchmarks for this configuration are still pending.
- Fine-tuning: While the model is pre-adapted for several systems, you will likely need to fine-tune it using specific datasets (e.g., the DROID research dataset) to achieve high accuracy on your specific robot arm or platform.
Verdict: Is it ready for production?
For Monday morning operations, Praxis-1 is not a 'set it and forget it' solution. Its inability to reliably plan complex, multi-step sequences means it currently functions best as an execution engine rather than a decision-making agent.
It excels at fast, reactive movements—such as grabbing objects or closing drawers—but requires a higher-level AI agent to break down complex workflows into individual, actionable steps. If you are building high-speed sorting or grasping systems, it is a powerful tool; if you need a robot to autonomously manage an entire warehouse floor, it is not there yet.
Sources
- Runway's open-source Praxis-1 turns video AI into a robot controller — the-decoder.com
Frequently asked questions
- Does Praxis-1 work autonomously?
- Not entirely. While it can execute movements based on visual input, it cannot reliably plan complex sequences of actions on its own and requires human-defined steps.
- What hardware do I need to run Praxis-1?
- The base model requires a GPU with 80 GB of VRAM. A compressed version is expected to run on 32 GB VRAM cards like the RTX 5090.
AI Briefing in 3 Minutes
Get what actually matters in AI & tech delivered free to your inbox. Zero spam, 1-click unsubscribe.
Comments
0 commentsDeixe seu comentário
Be the first to comment.
Continue Reading

Stop Scouting Creators Manually: Nectar’s New Agent Does It For You
Nectar launches AI agent for creator discovery

United Airlines Just Blocked AI Agents: What Now?
United Airlines bans autonomous AI booking agents

Google Just Dropped Gemini 4 Carbon — Here is the Catch With Swarm AI
Google releases Gemini 4 Carbon with swarm AI