Skild AI has introduced S1, a robot foundation model designed to learn previously unseen, long-horizon tasks from a single video demonstration without task-specific retraining. NVIDIA detailed the system on September 10 as part of a broader collaboration with Skild spanning model training, simulation, synthetic data and real-world deployment. Rather than updating its weights for every new task, S1 uses the demonstration directly as context to determine what the operator wants the robot to do.
The approach is based on in-context learning, a technique Skild compares with prompting in large language models. An operator records a demonstration of the desired task and provides the video to S1, which interprets the objects involved, the intended outcome and the sequence of actions before mapping them onto a physical robot. Skild says this allows one set of model weights to tackle tasks that were not included in its pretraining data, avoiding a separate fine-tuning process whenever a robot needs to learn something new.
Skild has demonstrated S1 completing tasks lasting as long as 10 minutes and involving dozens of manipulation steps. Examples include potting a plant, making pancakes, preparing pour-over coffee and assembling a kit, with the company saying the model had not encountered those exact tasks during pretraining. S1 is also intended to cope with differences between the demonstration and the real environment, including moved objects, unexpected disturbances and mistakes made during execution.
In one plant-potting test, Skild says the team progressed from recording the demonstration to autonomous execution on physical hardware in 11 minutes. The model was required to understand the demonstrated sequence rather than simply replaying a predefined set of movements. According to Skild, S1 can also recover from errors and combine previously learned manipulation abilities in new sequences when completing unfamiliar jobs.
The company has published internal benchmarks showing a substantial improvement over a comparable language-prompted system. On unseen multistep tasks after training with 100,000 hours of data, Skild says S1 achieved approximately 66% success at each step, compared with 9% for the alternative system. Skild further estimates that a single video demonstration can provide roughly the same benefit as about 380 conventional hands-on training examples, which the company says could otherwise require between 50 and 100 hours of manual data collection.
Those figures come from Skild’s own testing and have not been independently validated, but they illustrate the problem the company is attempting to solve. Industrial robots are typically optimized around stable processes, with changes to products, layouts or workflows potentially requiring more data collection, programming and validation. A system capable of learning a new workflow by watching it once could make robots considerably easier to redeploy across manufacturing and logistics environments where tasks change frequently.
NVIDIA technology is used throughout S1’s development pipeline. Skild trains its broader robot intelligence using NVIDIA accelerated computing alongside data from simulation, human videos, teleoperation and permitted commercial deployments. NVIDIA Cosmos tools are used to process and diversify video training data, while Omniverse and Isaac Sim provide simulated environments for generating data, evaluating edge cases and validating robot behavior before deployment.
Skild also uses NVIDIA Isaac Lab for reinforcement learning, with the Newton physics engine helping model properties such as contact, forces, collisions and pressure. NVIDIA and Skild are jointly developing additional GPU-accelerated simulation solvers intended to improve how robotic systems model gripping and physical interaction with solid objects. TensorRT is then used to optimize inference when models move onto physical hardware, where low response latency becomes important.
The technology is already being tested beyond research demonstrations. Skild, NVIDIA and Foxconn are deploying Skild’s robot intelligence on dual-arm manipulators used for precision assembly of NVIDIA Blackwell systems. One demonstrated workflow requires a robot to install a busbar and limit block, fasten 16 screws and continue operating when the physical scene changes, combining precise manipulation with sequence tracking and error recovery.
Skild says it has built more than 60 deployment partnerships covering manufacturing, logistics, inspection, security, food preparation and other applications, and NVIDIA says the company reached a $100 million annual revenue run rate within 10 months of its first commercial deployment. S1’s broader significance will depend on whether its one-video learning capabilities remain reliable across increasingly varied real-world environments. For now, the model represents an attempt to replace lengthy task-specific robot training with something much closer to showing a machine what to do and allowing it to work out the execution itself.
