Skild's S1: robots leave the BERT era — one video prompt, zero fine-tuning
deepakpathak · x · 2026-08-26
Skild AI founder Deepak Pathak introduces S1: in-context learning for robotics via the official blog. Core thesisLanguage modeling offers a blueprint: BERT-style models needed data collection plus fine-tuning per application, while ChatGPT's pivotal leap was in-context learning — injecting new concepts via prompt without touching weights. Robotics remains stuck in the BERT era: robust new-task execution still demands tens to hundreds of hours of post-training data in deployment conditions, and research (Oh et al., 2026) shows that with sufficiently dense post-training data, from-scratch policies match post-trained foundation models — so what is pre-training for?S1 capabilities- Unseen tasks never in pre-training - 10-minute horizons - Single video prompt, no post-trainingResponding to skeptics, Pathak likens it to a natural extension of system prompts as in Claude/GPT, and notes the blog shows the robot generalizes quite far beyond the prompt. He also teases more real user cases and, eventually, home robots.
More from Embodied
- Niantic Spatial's places library ships robot-ready Gaussian splat warehouse scenes for policy training — RexDouglass · 2026-10-16
- Volvo and NVIDIA expand partnership to bring safe automated driving to next-gen cars — nvidia · 2026-10-16
- Tevel's robotic system harvests fruit around the clock using computer vision — LexiLove · 2026-10-16
- SHOES: 10 minutes of real-robot RL lets quadrupeds walk in shoes, CoRL 2026 paper — kscottz · 2026-10-16
- Tesla Cybercab robotaxis logged zero crashes in first two weeks of public rides, NHTSA data shows — elonmusk · 2026-10-16
- Robotics LLMs tested on 16 tasks: GPT-6 Astra hits 17.2% success vs Fable's 9.1% — pabbeel · 2026-10-16