Skild AI Unveils Robot Foundation Model S1 with In-Context Learning

In official demonstrations, S1 completed long-horizon tasks such as making pancakes, brewing coffee, repotting plants, and assembling equipment. On unseen tasks, it achieved a 66% success rate, far exceeding the 9% achieved by language-prompted vision-language-action (VLA) models. To match S1's one-shot performance, traditional VLA models required approximately 380 task demonstrations through post-training.

Skild AI positions S1 as a shift from the "BERT era" to the "GPT era" for robotics. Just as large language models moved from task-specific fine-tuning to in-context learning, S1 uses video prompts instead of language prompts to learn new skills on the fly. The company argues that a true foundation model should not require hundreds or thousands of real-world demonstrations for every new task.

According to Skild AI's technical blog, the key test of in-context learning is whether the model can acquire novel skills never seen in training and recombine known skills into long, complex tasks. In the plant repotting task, the robot had to track its progress, decide next steps, and recover from mistakes, demonstrating true task understanding rather than simple behavior cloning. The entire process from recording the demonstration to robot execution took only 11 minutes.

The experimental results showed that with 100 hours of training data, language prompts outperformed ICL, but as data scale increased to 100,000 hours, ICL reached 96% on seen tasks versus 89% for language prompts, and 66% versus 9% on out-of-distribution tasks. Skild also found that language-prompt VLA performance degraded up to three times more under varying conditions, indicating that ICL learns to re-plan based on the current environment rather than repeating fixed actions.

Skild AI was founded in 2023 by CMU robotics professors Deepak Pathak and Abhinav Gupta. The company has raised $1.4 billion in a Series C round led by SoftBank, with participation from Nvidia and Bezos Expeditions, valuing it at over $14 billion. Skild emphasizes a cross-embodiment approach, aiming to deploy the same "Skild Brain" across robot arms, quadrupeds, and humanoids.