Skild AI's S1 Wants to Be the ChatGPT of Robots
A new foundation model lets robots learn complex, multi-minute tasks from a single video—no retraining required. But the ChatGPT moment hasn't arrived yet.
By The Robotics Desk
Skild AI has unveiled S1, a robot foundation model that lets machines pick up new tasks by watching a single video of a human performing them—no additional training required. The startup, founded in 2023, has raised nearly $1.7 billion to build what it calls a general-purpose robot brain, and S1 is its most public proof of concept yet. [1]
The capability is called in-context learning. A user drops a video of, say, a person repotting a plant or flipping a pancake into the model's prompt, and the robot attempts the task on its own. "You just add a video of a human doing something in the prompt, also called the context of the model, and it can just follow it on the robot," said Deepak Pathak, Skild AI co-founder and CEO. [1]
What distinguishes S1 from earlier robot AI isn't just what it can do—it's what it no longer needs. Typically, when faced with new tasks, AI models on robots need to be post-trained to handle them, a process that slows scaling and limits flexibility. S1 sidesteps that bottleneck by drawing on four distinct data streams during pretraining: teleoperation recordings, human videos, simulation data, and input from data-capture gloves worn by human demonstrators. [1]
Each data type carries its own trade-offs. Teleoperation is high quality—because the data comes directly from a robot—but slow to capture and doesn't provide a diverse range of data. Human video is abundant but hard to translate directly to robot hardware. Simulation scales cheaply but doesn't perfectly mirror the physical world. Data-capture gloves are slightly more scalable than teleoperation, but also slightly less directly applicable to robots. "If you look at any company out there, they are primarily focusing on one of these sources. But, if we think from first principles, we realize there is no golden path," Pathak said. "We have to use all of them because the pros of one source compensates for the downside of another source." [1]
This blended approach is the structural argument beneath the technical one. A model trained on a single data diet will always inherit its blind spots; mixing sources is less an engineering preference than a philosophical stance on how generalisation actually gets built.
The tasks Skild is showcasing are deliberately unglamorous and long: repotting a plant, making a cup of coffee, or cooking pancakes—activities that can take up to ten minutes. The pancake demo yielded a striking observation. "When the robot flipped the pancake, we were so paranoid that we went back and checked all of our millions of hours of data to look for any flipping example of any kind, and there was none," Pathak said. "This flipping basically emerged by looking at how the spatula is moving and the robot could just do it." [1]
S1 is also designed to be what Pathak calls "omni-bodied," meaning the same model can run on a quadruped, a static arm, or a humanoid robot. The company does plan to prioritise humanoid performance in future development, though it acknowledges the current version hasn't yet been fully scaled to that form factor. [1]
The analogy Pathak draws is the shift from task-specific language models to ChatGPT—when prompting replaced fine-tuning and general instruction-following became the norm. "The big change in language came when you had this transition to ChatGPT, where you don't have to fine-tune or train the network. You can just write in the prompt, and the model can follow it," he said. He doesn't think robotics is quite at its ChatGPT moment. "Now, the question is: Is it completely ready to be rolled out to people's homes? Not quite. But this is the first sign of what we believe might come." [1]
On the deployment side, Skild recently acquired Fetch Robotics—a move Pathak described as being about getting the right talent to be able to deploy and scale. He said S1 is already helping the company acquire customers and that more announcements are coming. "Of course, we'll continue scaling the frontier research, but at the same time, deployment is a high priority," Pathak said. [1]
The distinction between a compelling demo and a deployable product is where most robot AI companies have historically stalled. Skild's decision to name that gap openly, rather than paper over it, is either disarming honesty or a calculated way to manage expectations ahead of a commercial push—possibly both.
Foundation models reshaped software by removing the need to retrain for every new task. If S1's in-context approach holds up outside the lab, the same logic could start to apply to anything with a body. [1]
This flipping basically emerged by looking at how the spatula is moving and the robot could just do it.Deepak Pathak