Skild AI Launches S-1 Robot Foundation Model for Single-Video Task Learning

Skild AI has launched S-1, a robot foundation model that enables robots to learn previously unseen, long-horizon tasks from a single video demonstration, eliminating the need for new datasets, retraining, or task-specific post-training. For broader context, explore our AI News. For broader context, explore our Top 100 AI Tools.

S-1's In-Context Learning Capability

The S-1 model utilizes in-context learning to interpret demonstrated intent, objects, and sequences, then translates these into actionable commands for a robot. This approach allows the model to perform unfamiliar tasks lasting up to 10 minutes without requiring new datasets, retraining, or task-specific post-training. Examples of tasks S-1 can learn include plant potting, pancake making, pour-over coffee brewing, and kit assembly.

In a plant-potting test conducted by the Skild team, the process from recording a demonstration to autonomous execution on hardware was completed in 11 minutes. Skild estimates that a single short video example can be as effective as approximately 380 hands-on training examples, which would typically require 50 to 100 hours of manual data collection.

Performance Metrics and Efficiency

In Skild's internal tests on new multistep tasks, S-1 achieved a per-step success rate of approximately 66%. This compares to a 9% success rate for a similar AI system, highlighting S-1's efficiency in learning and executing new procedures.

Commercial Deployment and Partnerships

Skild AI has demonstrated rapid commercial growth, reaching a $100 million annual revenue run rate within 10 months of its first commercial deployment. The company has established over 60 deployment partnerships across various sectors.

A notable collaboration involves Skild, NVIDIA, and Foxconn, who are deploying the Skild Brain on dual-arm manipulators. This system is being used for high-precision assembly of NVIDIA Blackwell systems, including tasks such as installing 16 screws.

Underlying Infrastructure and Research

The development and research of S-1 were conducted on NVIDIA AI infrastructure. This includes tools and platforms such as Isaac Lab, Isaac Sim, Omniverse, and Cosmos world foundation models, which provided the environment for building and testing the advanced robotics model.

Conclusion

Skild AI's S-1 robot foundation model represents an advancement in robotic learning, offering a method for robots to acquire new skills from minimal input. Its ability to learn from a single video demonstration without extensive retraining could streamline the deployment of robots in diverse applications, potentially impacting industries requiring adaptable automation. The ongoing deployments with partners like NVIDIA and Foxconn indicate a practical application of this technology in complex manufacturing processes.

Sources

• Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video | NVIDIA Blog
• Skild AI S1: the robot brain that learns from one video | eesel AI
• Skild AI introduces S1, a foundation model that learns tasks...
• Skild AI's S1 Robot Model Combines Four Data Sources | James McNeil posted on the topic | LinkedIn