Details
- Skild AI has deployed its S1 robotic foundation model, which learns previously unseen, long-horizon physical tasks from a single video demonstration using in-context learning, without weight updates or task-specific post-training.
- The effort is part of a broad collaboration between Skild AI and NVIDIA, spanning synthetic data generation, large-scale training on NVIDIA accelerated infrastructure, simulation, and real-world deployment; partners include Foxconn for high-precision assembly of NVIDIA Blackwell systems.
- S1 lets operators record a video of a desired task—such as plant potting, pancake making, pour-over coffee, or kit assembly—and then maps the demonstrated intent, objects, and sequence into robot actions in new environments, adapting to disturbances and recovering from errors.
- In internal tests on new multistep tasks, S1 achieved about a 66% per-step success rate versus 9% for a similar AI system, and Skild estimates that one short video example can be as useful as roughly 380 hands-on training examples, compressing what could otherwise take 50–100 hours of data collection.
- Skild reports reaching a $100 million annual revenue run rate just 10 months after its first commercial deployment, with more than 60 deployment partnerships across manufacturing, logistics, inspection, security, food preparation and other applications, all built on NVIDIA technologies such as Cosmos, Omniverse, Isaac Sim, Isaac Lab, Nsight, TensorRT and upcoming Newton solvers.
Impact
S1 and the NVIDIA–Skild collaboration point to a shift in industrial robotics from static, preprogrammed automation toward adaptable, foundation-model-driven physical AI that can be re-tasked via simple video prompts. Over the next 12–24 months, this approach is likely to accelerate deployment of generalist robot brains in factories and warehouses, influence R&D toward unified data-simulation-training pipelines, and draw more funding to in-context, multimodal robotics platforms competing with other foundation-model strategies.
