Skild AI launches robot model that learns new tasks from a single video
Learning from context without retraining
Skild AI released S1 last week, a robotics foundation model that can learn long-horizon tasks it has never seen from a single video demonstration — no weight updates, no dedicated fine-tuning. The technique, called in-context learning, lets an operator record the desired task, feed the video as a prompt, and have the model interpret the intent, objects and action sequence directly into commands for the robot in front of it. The model was built on NVIDIA's AI infrastructure under a broad collaboration covering synthetic data generation, training, simulation and physical deployment.
Performance that speaks in numbers
In internal tests on novel multi-step tasks, S1 succeeded on roughly 66% of steps per attempt, versus 9% for a comparable system — a more than sevenfold improvement. Skild estimates a single short demonstration video equals about 380 manual training examples, which would take 50 to 100 human hours to collect. In a plant-potting experiment, the team went from recording the demo to autonomous execution on hardware in just 11 minutes. The model also adapts when objects move, recovers from errors and composes skill sequences that were never explicitly programmed.
Tasks up to ten minutes with no new dataset
S1 performs unfamiliar tasks lasting up to ten minutes — potting plants, making pancakes, brewing pour-over coffee, assembling kits — each requiring dozens of manipulation steps and skill compositions the model had never executed before. The approach breaks the cycle in which every new product, process or factory deployment demands fresh data collection, retraining and re-validation. Where customer agreements allow, experience from commercial deployments feeds back into the shared model and accelerates future rollouts.
Hundred-million-dollar revenue run rate in ten months
The launch arrives as the company hit a $100 million annual revenue run rate just ten months after its first commercial deployment. In that span Skild has built more than 60 deployment partnerships across manufacturing, logistics, inspection, security, food preparation and other applications. "The shift from pre-programmed learning to learning from experience is the real step change in robotics," said Deepak Pathak, Skild co-founder and CEO, "and NVIDIA's Isaac Lab and Cosmos technologies provide the diverse, scalable experience robots need."
On the Foxconn factory floor
The work is already live on the production line: Skild, NVIDIA and Foxconn are deploying "Skild Brain" on dual-arm manipulators for precision assembly of NVIDIA's Blackwell systems. In a demonstrated workflow, the robot installs a busbar and retention block, tightens 16 screws and adapts to perturbations across a multi-step task — a capability that demands precise motion, touch-aware control, sequence tracking and recovery when the scene deviates from plan.
Accelerated infrastructure across the development cycle
NVIDIA accelerated compute gives Skild the scale to train the shared robot brain through simulation, human data and physical deployment, all in a single pipeline that shortens the path from lab to factory.