Skild's S1 learns pancake flipping from one video, no fine-tuning, as robot ICL emerges as a trend
deepakpathak · x · 2026-09-09
- Skild AI's general-purpose robotics foundation model S1 watched a human flip a pancake once, then did it on real hardware—even though pancake flipping wasn't in its pre-training data.
- With the same weights and no fine-tuning or task-specific post-training, it handles unseen tasks like potting plants, pour-over coffee and kit assembly, with multi-step runs up to 10 minutes.
- Potting a plant took just 11 minutes from human demo to autonomous execution; at 100K hours of pre-training, S1 hit 66% average per-step success on unseen tasks.
- NVIDIA, Skild, Generalist and X Square have all shared demo-learning work in recent weeks—robot in-context learning is becoming a trend that could cut the collect-finetune-deploy loop.
Related event: Skild's robot foundation model S1: one .mp4 is the robot's .exe(7 posts)→
More from Embodied
- Munro Live Posts Full Cybercab Engineering Deep-Dive Video — yunta_tsai · 2026-10-27
- Tesla's Robotaxi Fleet Keeps Growing in Austin — yunta_tsai · 2026-10-27
- Building Halo Infinite Forge maps with local LLMs, DCjson and Blender MCP — swagonflyyyy · 2026-10-27
- SoftBank-Backed XAG Ships Drones With Autonomous Docks That Clean, Recharge and Refill — pstAsiatech · 2026-10-27
- Toyota Enters Humanoid Robot Race Against Tesla and China With 20 Years of Research — pstAsiatech · 2026-10-27
- Robot learning researcher joins Mila to work on physical AI and RL — pcastr · 2026-10-27