Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video

Skild AI has introduced a new S1 robot foundation model that can learn new tasks from a single video demonstration. This breakthrough allows robots to adapt to changing manufacturing floors, warehouses, and production lines without significant reprogramming. By using in-context learning, S1 can understand and execute tasks without updating its weights or undergoing task-specific post-training.

The model was developed using NVIDIA AI infrastructure and is part of a collaboration to move adaptable robot intelligence into dynamic operating environments. Skild's CEO, Deepak Pathak, emphasized that learning by experience rather than preprogramming marks a significant step in robotics. The S1 model has shown impressive results, achieving about 66% success in each step for new tasks, compared to 9% for a similar AI system.

This success rate is equivalent to the robot gaining experience from around 380 hands-on training examples, which would typically take a person 50-100 hours to collect manually. By allowing operators to demonstrate new tasks directly, S1 breaks the cycle of needing constant retraining for every change. Skild, NVIDIA, and Foxconn are already deploying the Skild Brain on dual-arm manipulators for high-precision assembly of NVIDIA Blackwell systems, demonstrating the robot's ability to install components, fasten screws, and adapt to disturbances.

NVIDIA's accelerated computing, open world foundation models, and simulation frameworks play a crucial role in training and validating Skild's robot brain before deployment. The collaboration between Skild AI and NVIDIA aims to streamline the development cycle, connecting data, simulation, training, and deployment stages using NVIDIA's technologies.