Skild's S1 learns new robot tasks from a single demo video, hitting 66% success vs 9% for VLA
量子位 · wechat · 2026-08-26
Embodied-AI startup Skild AI unveiled S1, a robot foundation model built around in-context learning (ICL): with zero post-training, it performs unseen long-horizon manipulations after watching a single human demo video, extending ICL task length beyond 10 minutes. Demos include pancake flipping, coffee brewing, plant repotting (dozens of sequential steps), and device assembly; the repotting task went from demo recording to robot execution in 11 minutes, with frozen model weights throughout.
Key numbers: On OOD tasks, ICL hits 66% success versus 9% for language-prompted VLA; a conventional VLA needs roughly 380 post-training demos to match one-shot S1 (86% at 2,000 demos). As training data scaled from 1,000 to 100,000 hours, ICL overtook language prompting — Skild's proposed "ICL scaling law": more data makes the model better at learning from demonstrations. Language-prompt VLA degrades up to 3× more under distribution shift.
Skild argues that a foundation model requiring tens to hundreds of hours of real-robot data per new task is no foundation model at all — robotics is moving from the BERT era (per-task fine-tuning) toward a GPT-3 moment (learn from examples). Founded in 2023 by CMU professors Deepak Pathak and Abhinav Gupta, the company champions "Any robot, any task, one brain" across embodiments and uses teleop, UMI, egocentric video, and simulation data alike. It raised a $300M Series A at a $1.5B valuation in 2024, then a $1.4B Series C led by SoftBank in January at over $14B. Observers say the robot GPT moment may be arriving, with generalist robots possibly within two years.
More from Embodied
- Viral meme nails AI's irony: it can write symphonies but can't do the dishes — briannekimmel · 2026-10-16
- WEDGE turns navigation agents' own failures into targeted training worlds — mohitban47 · 2026-10-16
- Knödel Watch: animated vocabulary cards on ESP32-S3, open-sourced 4 days after unboxing — alexcovo_eth · 2026-10-16
- NVIDIA RTX Spark PC launches tomorrow with global events and strong early reviews — NVIDIA Blog · 2026-10-16
- SlipSense fuses pressure and vibration for 23ms zero-shot robotic slip detection — micoolcho · 2026-10-16
- Stanford's Fei-Fei Li team open-sources OpenWAM, a unified world-action model framework for robots — jiqizhixin · 2026-10-16