Skild AI has released S1, a robot foundation model the company says can carry out manipulation tasks it never encountered during training after watching a single video demonstration. The sequences run as long as ten minutes.

 

The approach breaks with how most robot policies are built today. Instead of collecting hundreds of demonstrations and fine-tuning a model for each new job, S1 treats the demonstration itself as a prompt and executes with unchanged weights.

 

Skild published the model and its evaluation results on August 25, and trade coverage continued through the first week of September as the robotics industry worked through the claims.

 

Further Reading

 

How S1 Reads a Video Prompt

 

S1 was built from the start as an in-context learner. During pre-training it is fed episodic data in which the task is specified only visually, forcing the model to infer the demonstrator's intent and track how far along the task has progressed.

 

The practical effect is that a new task requires no gradient updates at all. The same weights that potted a plant are the weights that make a pour-over coffee, with only the prompt changing.

 

"You just add a video of a human doing something in the prompt, and it can just follow it on the robot."

 

The description came from Skild chief executive Deepak Pathak, who co-founded the company in 2023.

 

Skild's demonstrations cover pancake cooking, pour-over coffee, plant repotting and mechanical kit assembly. The company says the pancake-flipping behaviour was not represented in its training examples and emerged from the model.

 

Skild's Success-Rate Comparison

 

The headline figure is a comparison against a conventional vision-language-action baseline. On tasks neither system had seen, S1 reached a 66 percent step-level success rate at 100,000 hours of pre-training data, against 9 percent for the language-prompted model.

 

Skild also puts a conversion rate on the demonstration itself. One video supplied in context is worth roughly 380 post-training examples, a volume that would take 50 to 100 hours of teleoperation to gather.

 

The company is candid that brute force still wins on a task you already know about. A vision-language-action model given 2,000 demonstrations of a specific task reaches 86 percent, comfortably ahead of S1's in-context result.

 

That trade-off is the actual argument. S1 is not claiming to beat a heavily post-trained specialist, but to remove the weeks of data collection standing between a robot and a task nobody anticipated.

 

Four Data Streams Behind the Model

 

Skild's pre-training mix pulls from four sources, each chosen to cover another's weakness:

 

  • Robot teleoperation, which is high quality but slow and expensive to capture
  • Egocentric human video, which is abundant and diverse but hard to map onto hardware
  • Universal Manipulation Interface data-capture gloves, which scale moderately well
  • Simulation, which scales without limit but carries a reality gap

 

Pathak has argued that rivals tend to concentrate on one of these streams and inherit its specific failure mode. Skild's position is that blending all four cancels out the individual weaknesses.

 

The company also says it spends three dollars on quality control for every dollar spent collecting data, an unusually high ratio for a field that has generally competed on raw hours. S1 runs on Nvidia AI infrastructure.

 

Emergent Behaviour and Where It Breaks

 

Skild reports several behaviours it did not explicitly train. The model recovers from its own errors, tolerates objects being moved mid-task, and will substitute a different object when the one in the demonstration is missing.

 

More unusually, S1 sometimes outperforms the video it was given. Because it treats the demonstration as a statement of the goal rather than a trajectory to copy, it can smooth out a clumsy human motion instead of reproducing it.

 

The honest caveats are visible in Skild's own numbers. A 66 percent step-level success rate means a ten-step task fails more often than it finishes, and step-level scoring is a more forgiving metric than end-to-end completion.

 

Independent replication is also absent. Every result published so far is Skild's own, on Skild's task set, and the robotics field has no shared benchmark comparable to the leaderboards that discipline language models.

 

Warehouse Assets and the Humanoid Gap

 

Skild describes S1 as omni-bodied, running across quadrupeds, static robotic arms and humanoids from the same checkpoint. Humanoid-specific optimisation is a stated future priority rather than a shipped capability.

 

The commercial route runs through hardware the company bought rather than built. Skild acquired Zebra Technologies' robotics automation business, the assets that were formerly Fetch Robotics, giving it an installed base of warehouse machines to deploy against.

 

Skild raised $1.4 billion in a SoftBank-led round in January that valued it above $14 billion, with Nvidia also participating. Total funding stands near $1.7 billion since the company was founded three years ago.

 

The framing Skild uses for the wider field is that robotics is stuck in a "BERT era" of pre-train-then-post-train, waiting for the shift that in-context learning brought to language models. S1 is the company's argument that the shift has started.

 

Whether it holds depends on results outside Skild's own lab. Until a customer deployment or an independent evaluation puts the single-video claim under pressure, S1 is a striking demonstration rather than a settled capability.