Skild AI has launched its S1 robot foundation model, leveraging NVIDIA AI infrastructure to execute long-horizon, multistep tasks from a single video demonstration. Operating via in-context learning without updating model weights, the system achieves a 66% step-by-step success rate in dynamic environments like manufacturing floors and warehouses.
Beyond Fixed Programming: How the S1 Model Interprets Video Demonstrations
Industrial automation has long been bottlenecked by rigid code. Manufacturing floors, warehouses, and production lines shift constantly as new layouts, products, and processes arrive. Traditional industrial robots require deep data collection, retraining, and validation for every minor adjustment. Skild AI bypasses this costly cycle with its S1 robot foundation model, launched in September 2026. The system utilizes in-context learning to interpret video prompts supplied by human operators.
Instead of undergoing task-specific post-training or adjusting its neural network weights, the S1 model takes a short video of a desired task and immediately maps the demonstrated intent, objects, and sequence into direct physical actions. In benchmark testing highlighted by Skild AI, the S1 robot successfully executed unfamiliar tasks lasting up to 10 minutes. These operations spanned plant potting, pancake making, pour-over coffee brewing, and kit assembly, requiring the robot to compose complex manipulation sequences it had never encountered in its pretraining dataset.
Quantifying the Leap in Robotic Adaptability
The operational efficiency gains of video-prompted in-context learning are stark when measured against legacy automation approaches. In controlled multistep evaluations conducted by Skild AI, the S1 robot achieved a success rate of approximately 66% at each individual step. By comparison, a similar AI system managed only a 9% success rate under the same conditions, marking a more than sevenfold performance improvement.
Furthermore, Skild AI estimates that providing a robot with a single short video demonstration yields a training utility roughly equivalent to 380 hands-on training examples. Manually collecting that volume of physical training data typically demands between 50 and 100 hours of human labor. In practical plant-potting tests, Skild AI engineers transitioned from recording a human video demonstration to autonomous execution on physical hardware in just 11 minutes.
The model’s underlying mechanics allow it to adjust dynamically when objects shift position, autonomously recover from execution errors, and sequence manipulation skills that were never explicitly programmed into the machine.
The NVIDIA Infrastructure Stack Driving Physical AI
Building a foundation model capable of generalizing across varied robot embodiments requires massive computational scale. Skild AI developed the S1 model and conducted its core research on NVIDIA AI infrastructure. This collaboration spans synthetic data generation, simulation, model training, and real-world deployment.
“Learning by experience, and not preprogramming, is the step change that has happened in robotics,” said Deepak Pathak, cofounder and CEO of Skild AI, in official statements released by NVIDIA. “NVIDIA Isaac Lab and NVIDIA Cosmos technologies help Skild create the scalable, diverse experience its robots need to learn across many scenarios and embodiments.”
The development pipeline relies on several core NVIDIA systems:
- NVIDIA Cosmos: Open world foundation models that diversify training data, turning raw video feeds into structured descriptions.
- Cosmos Curator: Software used to annotate, filter, and organize massive datasets at scale.
- NVIDIA Isaac Sim & Omniverse: Physically based virtual environments utilized to generate synthetic data, test edge cases, and validate behaviors prior to physical deployment.
- Isaac Lab & Newton Physics Engine: An open modular robot learning framework powered by the Newton physics engine, enabling engineers to model forces, contact, collisions, and pressure to narrow the simulation-to-reality gap.
- Nsight & TensorRT: Developer tools that locate performance bottlenecks during training and optimize inference so deployed robots can react swiftly in physical environments.
Commercial Traction and Factory Floor Integration
Skild AI has scaled rapidly since its commercial deployment, reaching a $100 million annual revenue run rate just 10 months after its initial rollout. The company has established more than 60 deployment partnerships spanning logistics, inspection, security, food preparation, and manufacturing.
The most demanding testbeds are already live. Skild AI, NVIDIA, and Foxconn are actively deploying the Skild Brain on dual-arm robotic manipulators for the high-precision assembly of NVIDIA Blackwell systems. In these factory floor workflows, a robot must install a busbar and limit block, fasten 16 screws, and maintain contact-aware control while recovering from physical disturbances.
Where customer agreements permit, operational experience gathered from these commercial deployments flows back into the broader model architecture. This feedback loop accelerates future deployments, bridging the gap between isolated laboratory research and resilient, scalable factory work.
