What happened
Skild AI launched its S1 robot foundation model, which uses in-context learning to execute previously unseen tasks from a single video prompt. The company reported reaching a $100 million annual revenue run rate and is deploying the technology with Foxconn for high-precision assembly.
Skild AI launched the S1 robot foundation model, designed to learn previously unseen, long-horizon tasks from a single video demonstration. The model utilizes in-context learning, meaning it interprets the demonstrated intent, objects, and sequence from the video prompt and maps them into actions without updating its weights or undergoing task-specific post-training.
The company stated that S1 can perform unfamiliar tasks lasting up to 10 minutes, such as plant potting, pancake making, and kit assembly. In one internal test, the team moved from recording a demonstration to autonomous execution on hardware in 11 minutes. Skild reported that S1 achieved a 66% success rate per step on new multistep tasks, compared to 9% for a similar AI system, and estimated that one video example is as useful as roughly 380 hands-on training examples.
Skild AI built and researched S1 on NVIDIA AI infrastructure, including NVIDIA Isaac Lab, Cosmos, and Omniverse. The collaboration spans synthetic data generation, model training, simulation, and real-world deployment. The company has reached a $100 million annual revenue run rate 10 months after its first commercial deployment and has established more than 60 deployment partnerships across manufacturing, logistics, and other sectors.
A specific deployment is underway with Foxconn, where the Skild Brain is being used on dual-arm manipulators for high-precision assembly of NVIDIA Blackwell systems. This workflow involves installing components, fastening screws, and adapting to disturbances, requiring precise motion and contact-aware control.
Source details: blogs.nvidia.com ↗
Why it matters
This development addresses a core limitation in industrial robotics: the high cost and time required to reprogram or retrain robots for new tasks or layouts. By enabling robots to learn from a single video demonstration, S1 significantly reduces the barrier to adapting automation in dynamic environments like manufacturing and logistics. This shift from fixed, preprogrammed workflows to adaptable, experience-based learning could accelerate the adoption of general-purpose robotics in industries where product lines and processes change frequently.
Industrial robots are traditionally built for fixed jobs, requiring significant data, retraining, and validation for each new product or process change. S1’s ability to learn from a single video demonstration breaks this cycle, allowing operators to demonstrate new tasks directly without creating new datasets or running training cycles for every change.
This capability is particularly relevant for dynamic operating environments like manufacturing floors and warehouses, where layouts shift and new products arrive frequently. By reducing the time and cost associated with robot adaptation, S1 may make it more feasible for companies to deploy flexible automation in response to changing production needs.
The integration with NVIDIA’s simulation and training tools, such as Isaac Lab and Cosmos, highlights a trend toward using synthetic data and physically based virtual environments to bridge the gap between simulation and real-world performance. This approach allows for extensive testing of edge cases and behaviors before physical deployment.
What to watch next
Monitor the real-world success rates of S1 in commercial deployments beyond the reported 66% per-step success rate. Watch for the release of the jointly developed GPU-accelerated simulation solvers to the broader developer community and the expansion of Skild’s 60+ deployment partnerships.
The reported 66% success rate per step in Skild’s tests indicates that while the model is significantly more effective than previous systems, it is not yet perfect. Real-world performance in complex, uncontrolled environments may vary, and monitoring the reliability of these deployments will be crucial for assessing the technology’s practical viability.
Skild and NVIDIA are jointly developing new GPU-accelerated simulation solvers for modeling physical interactions, which are planned to be made available to all developers as part of Newton. The release of these tools could lower the barrier to entry for other robotics companies looking to implement similar simulation-based training pipelines.
The expansion of Skild’s commercial partnerships, currently numbering over 60, will determine the breadth of the technology’s application. Observing how the model performs across different industries, such as food preparation and security, will provide insight into its generalizability beyond manufacturing.