Editorial illustration for Skild AI Trains Robots With Single Videos in 11 Minutes
Skild AI Trains Robots From Single Videos in 11 Min
Skild AI says its S1 robot foundation model can pick up a new task after watching a single video, no retraining required. The Pittsburgh startup launched the model last week, built and trained on NVIDIA AI infrastructure, and it targets a problem that has dogged industrial robotics for years: factory floors and warehouses change constantly, but most robots don't. A new product line or a shifted layout usually means weeks of reprogramming and fresh data collection before a machine can work again.
S1 skips that cycle through what Skild calls in-context learning. The robot watches a video demonstration and executes the task without updating its underlying weights or going through task-specific post-training. That's a departure from how industrial robots have typically been deployed, where every new job means starting the retraining process over.
The timing matters for Skild's business as much as its technology. The company hit a $100 million annual revenue run rate just 10 months after its first commercial deployment, and it now counts more than 60 deployment partnerships across manufacturing, logistics, inspection, security and food preparation.
S1 takes a different approach: An operator records a video of the desired task and provides it to the model as a prompt. It interprets the demonstrated intent, objects and sequence, then maps them into actions for the robot in front of it — with no retraining — and often for a task not covered by its pretraining dataset.
Why this matters
Eleven minutes from demo to autonomous execution is the number worth sitting with. If Skild's S1 genuinely generalizes from a single video without retraining, that changes the cost math for anyone deploying robots on lines that change weekly, not yearly. Founders building warehouse or manufacturing automation should watch whether this holds outside curated demos, since "new, unseen tasks" claims from robotics labs have a long history of shrinking once independent labs get hardware access.
Researchers should push Skild on what "long-horizon" and "recover from errors" actually cover, error recovery in a controlled plant-potting test is a different animal than a jammed conveyor at 2 a.m. NVIDIA's involvement through its Physical AI stack signals where the infrastructure money is flowing, and that's worth tracking regardless of how S1 itself performs. For now, the 11-minute figure is a strong marketing anchor.
Whether it's also an engineering reality across messier, real-world floors is the thing to verify before anyone reworks a deployment roadmap around it.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives Pocket FM Doubles Revenue Run Rate to USD 500M as AI Powers Audio Content
From Archives Partnership on AI Adds Six New Partners for Global Progress Hub
Common Questions Answered
How does Skild AI's S1 robot foundation model learn new tasks without retraining?
The S1 model learns new tasks by having an operator record a video of the desired task and provide it as a prompt to the model. The model interprets the demonstrated intent, objects, and sequence from the video, then maps them into actions for the robot without requiring any retraining or additional data collection.
What is the significance of the 11-minute timeframe mentioned for Skild AI's S1 robot?
The 11-minute timeframe represents the time it takes from recording a demonstration video to the robot autonomously executing the new task. This rapid deployment capability changes the cost economics for factories and warehouses that frequently change product lines or layouts, eliminating the weeks of reprogramming and data collection that traditional robots require.
What infrastructure does Skild AI use to build and train its S1 robot foundation model?
Skild AI built and trained its S1 robot foundation model on NVIDIA AI infrastructure. This partnership leverages NVIDIA's physical AI capabilities to enable the model's advanced video-to-action learning capabilities.
Why has adapting to changing factory environments been a persistent challenge in industrial robotics?
Most industrial robots are not designed to adapt quickly when factory floors and warehouses change, such as when new product lines are introduced or layouts are shifted. Traditionally, these changes require weeks of reprogramming and fresh data collection before a machine can work again, making it costly and time-consuming to update robot operations.
Can Skild AI's S1 model perform tasks that were not included in its pretraining dataset?
Yes, according to Skild AI's approach, the S1 model can often handle tasks that were not covered by its pretraining dataset. The model generalizes from the single video demonstration provided by the operator, allowing it to execute novel tasks beyond what it was originally trained on.
Further Reading
- Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video - NVIDIA Blog
- Introducing S1: In-Context Learning for Robotics - Skild AI
- Skild AI S1: Robot Model Learns Tasks From One Video - DataNorth AI
- Skild AI unveils S1 robotics model that learns physical tasks from a single video - Crypto Briefing
- Skild AI cofounder Gupta says robots hit ChatGPT moment, learn from one video - The Chosun Ilbo
