Skild AI Unveils S1, a Robot Foundation Model That Learns 10-Minute Tasks From a Single Video

On August 26, Skild AI released S1, a generalist robot foundation model built around in-context learning: given a single video demonstration and zero fine-tuning, it can execute novel, multi-step tasks lasting 10+ minutes that never appeared in its pretraining data, in real time. Multiple commentators called it a potential "GPT moment" for robotics.

Confirmed

• Founder Deepak Pathak's blog frames S1 as an in-context learner from the ground up, analogous to the leap from BERT-style per-task fine-tuning to ChatGPT's in-context learning (per @dioscuri relaying Andrew Curran).
• Tasks are specified via video rather than language prompts; S1 maps the video to its own scene, sidestepping the imprecision of language instructions, and extends this capability from locomotion to manipulation (@rohanpaulai).
• Demonstrations include making pancakes from one video demo, plus bed-making, coffee brewing, planting, kit assembly and pour-over coffee — all 10+ minutes with dozens of steps, with instruction-following and recovery from errors.
• Per @DeryaTR, internal benchmarks show a 66% step success rate on unseen tasks; a single video demo rivals roughly 380 language-prompted samples.
• Learning a new task previously required 100 hours of training data; S1 compresses this to one demo (@heyshrutimishra).
• Per @lukasmziegler, the team found no pancake-making footage in millions of hours of pretraining data — S1 inferred the task from the single demo.

Why it matters

• If single-shot, zero-fine-tuning generalization holds, robot skill development could shift from per-task engineering to demonstration-based learning — widely framed as the onset of generalist robot intelligence (@Scobleizer, @rohanpaulai, others).
• A comparison shared by @deepakpathak notes that Skild (S1), Generalist and Sunday (ACT-2, 99.1% shirt-folding in unseen homes but single-task) all shipped results within five weeks, giving a concrete yardstick for foundation-model progress; S1's differentiator is multi-task, long-horizon manipulation from one video.

2026-08-26 ~ 2026-08-27 · 40 related posts

• [source] Skild AI unveils S1: robot foundation model learns 10-minute novel tasks from a single video prompt — deepakpathak · 2026-08-26
• [source] Nobody taught it to flip pancakes: Skild's S1 learns new tasks from one demo — lukas_m_ziegler · 2026-08-26
• Skild robot learns to make pancakes from one video, demonstrating long-horizon task execution — Scobleizer · 2026-08-26
• SkildAI Unveils S1: One-Shot Learning Model for Robotics Tasks — deepakpathak · 2026-08-26
• SkildAI releases S1 robot model that learns unseen tasks from one video demo — deepakpathak · 2026-08-26
• Skild releases S1, a robot foundation model that learns tasks from video demos — deepakpathak · 2026-08-26
• [source] Skild's S1: robots leave the BERT era — one video prompt, zero fine-tuning — deepakpathak · 2026-08-26
• Robotics In-Context Learning: Execute 10-Minute Tasks with One Video Prompt, No Fine-Tuning — rohanpaul_ai · 2026-08-26
• Skild AI unveils S1: robots learn 10-minute novel tasks from a single video — Haoyu_Xiong_ · 2026-08-26
• Robot Learns New Tasks from Single Video Demo Without Fine-Tuning — dioscuri · 2026-08-26
• Skild AI Unveils S1: Robotic In-Context Learning with One Demo — CyberRobooo · 2026-08-26
• SkildAI Unveils S1: Learning Complex Tasks from One Example — chris_j_paxton · 2026-08-26
• Skild's S1 learns new robot tasks from a single demo video, hitting 66% success vs 9% for VLA — 量子位 · 2026-08-26
• Skild AI unveils S1: robot foundation model that learns 10-minute novel tasks from one video — deepakpathak · 2026-08-26
• SkildAI S1 achieves one-shot imitation from video — deepakpathak · 2026-08-26
• SkildAI robots handle 10+ minute tasks: pancakes, coffee, potting — deepakpathak · 2026-08-27
• Robotics Foundation Model Progress: Skild S1 Enables In-Context Learning — deepakpathak · 2026-08-27

23 near-duplicate retellings: deepakpathak · pratyushmaini · xennygrimmato_ · deepakpathak · deepakpathak · heyshrutimishra · chris_j_paxton · chris_j_paxton · deepakpathak · deepakpathak · deepakpathak · deepakpathak · DeryaTR_ · rohanpaul_ai · deepakpathak · deepakpathak · deepakpathak · deepakpathak · deepakpathak · deepakpathak · deepakpathak · deepakpathak · lukas_m_ziegler