Revolutionizing Robotics: Skild AI&#8217;s S1 Model Achieves ChatGPT-like Intelligence in the Physical World
<video style="width: 100%;height: auto" controls><source src="https://d2qu7nggdaeihr.cloudfront.net/output/5075321d-09ce-4de9-8c72-42f1a924b422-1794329025.mp4" type="video/mp4"></video>
<aside>(This article was generated with AI and it&#8217;s based on a AI-generated transcription of a real talk on stage. While we strive for accuracy, we encourage readers to verify important information.)</aside>
<img decoding="async" src="https://web-summit-engine.imgix.net/ddklk4wu5l8v43qxchi8ts9h3drz?auto=format&amp;fit=fill&amp;fill=solid&amp;fill-color=white&amp;w=600&amp;h=600&amp;fm=jpg&amp;dl=Abhinav Gupta.jpg" alt="Abhinav Gupta" align="left" hspace="10" width="200" style="float:left;margin-right: 15px">
The AI revolution, driven by vast data and compute, faces unique challenges in robotics. The physical world lacks readily available large-scale data, and hardware is often slow and expensive. This impedes rapid progress, preventing robotics from achieving the revolutionary scale seen in other AI domains.

Professor Abhinav Gupta, Co-founder and President of Skild AI, introduced their solution: the world&#8217;s first scalable robotics foundation model. Skild AI&#8217;s model is trained on 100 times more data than competitors. It boasts multi-scenario, multi-task, and multi-hardware capabilities, enabling one model to operate across diverse environments, perform various functions, and work with different robot types.

Skild AI&#8217;s vision is &#8220;omnibodied intelligence,&#8221; aiming to replicate the ChatGPT movement for robotics with the motto: &#8220;Any robot, any task, one brain.&#8221; ChatGPT&#8217;s breakthrough was &#8220;in-context learning,&#8221; allowing it to grasp new problems from a single example. Conventional robotics, however, demands extensive post-training, often requiring 50-100 hours of data collection and fine-tuning, significantly slowing adoption.

To overcome this, Skild AI launched its S1 model in August, a general-purpose robotics foundation model that transforms physical AI. S1 enables robots to learn a wide range of new, complex, and long-horizon tasks (over ten minutes) from just one video demonstration. This in-context learning allows robots to perform actions like making pancakes or coffee, even if these specific tasks were not part of their initial training data.

S1&#8217;s generalization capability stems from scaling pre-training by combining human videos, which convey task objectives, with simulation, providing practice under varied conditions. This, alongside high-quality deployment data, established the first scaling law of robotics. S1 performs comparably on known tasks but significantly excels on unseen generalization tasks, demonstrating robust adaptability.

The model&#8217;s in-context learning also facilitates dynamic adaptation to hardware changes, such as a robot continuing to function effectively despite a broken leg. Skild AI emphasizes practical, real-world deployments, having achieved a $100 million annual revenue run rate in August, serving over 60 paying customers within ten months across diverse sectors including manufacturing, logistics, and inspection.

Key deployments include automating GPU installation at Nvidia&#8217;s Houston factory, where a Skild AI robot achieved 62 units per hour (exceeding human speed) with 100% accuracy, utilizing only three hours of deployment data. The versatile S1 model also handles tasks like cooking eggs or packing AirPods. Skild AI robots also conduct inspections in industrial sites, data centers, and airports, autonomously generating facility tickets.

Looking forward, Skild AI plans to integrate self-improvement capabilities, allowing deployed models to enhance their performance on the job through reinforcement learning and self-play. An experimental demonstration showcased robots learning to play football through 400 years of simulated self-play, illustrating a future where a single AI brain learns extensively, deploys efficiently, and continuously refines its abilities.

When AI leaves the screen

AI: doom or boom?

What comes after the Blue Wave?