Back to feed
News Story
量子位
1 sources

Skild AI Unveils Robot Foundation Model S1: Learns Long Tasks from a Single Demonstration Without Fine-Tuning

Skild AI has released a new robot foundation model, S1, that leverages in-context learning to perform complex tasks up to 10 minutes long by watching a single human demonstration, without any fine-tuning. It achieves a 66% success rate on unseen tasks, far surpassing language-prompted VLA models at 9%. This marks a potential paradigm shift in robotics, moving from a BERT-like era to a GPT-like era, which could dramatically reduce the data cost for developing robot skills.

SynthePulse Insight · AI deep readingMembers

Skild AI Unveils S1: A Breakthrough in Robot In-Context Learning

Version 1 · 1 source

Skild AI introduces S1, a robot foundation model that uses in-context learning to let robots complete complex tasks up to 10 minutes long after watching a single human demonstration, with a 66% success rate—far surpassing the 9% of traditional VLAs.

  • Skild AI releases robot foundation model S1, supporting in-context learning tasks up to 10 minutes long.
  • S1 achieves a 66% success rate on unseen tasks, while language-prompted VLAs only reach 9%.
  • Traditional post-training requires about 380 demonstrations to match S1's one-shot performance.
Open section navigationS1's Core Capability: In-Context Learning

S1's Core Capability: In-Context Learning

Skild AI's robot foundation model S1 focuses on in-context learning (ICL), enabling robots to learn new tasks without post-training, simply by watching a single human demonstration video. In official demos, the robot completed long-horizon tasks such as making pancakes, brewing coffee, repotting plants, and assembling equipment, with task lengths up to 10 minutes.

S1 achieves a 66% success rate on unseen tasks, far surpassing language-prompted VLAs (9%). Traditional post-training requires about 380 demonstrations to match S1's one-shot performance, and in the plant repotting task, S1 went from recording the demo to robot execution in just 11 minutes.

Throughout the process, S1 used no fine-tuning or post-training; the model weights remained completely unchanged, using the same set of weights for all demonstrated tasks.

Free for now

Read the full analysis

4 more sections of analysis, plus the full takeaway

Loading

Credibility boundary

This article primarily draws from QbitAI's report, which includes citations from Skild AI's official blog. All data (such as success rates and demonstration counts) come from Skild AI's official statements or demos and have not been independently verified. Some industry comments (e.g., netizen opinions) are personal views and do not represent facts.

Sources for this version

  1. 硅谷今日最热具身模型！不用后训练，看一遍就学会量子位

Primary report

量子位

Primary source