Just ten months after its first commercial deployment, Skild AI has reached a $100 million annual revenue run rate while establishing over 60 partnerships across manufacturing, logistics, and even food preparation. The company’s new S1 robot foundation model achieves adaptability by learning previously unseen tasks from a single video demonstration, eliminating the need for reprogramming, Skild AI says.

“Learning by experience, and not preprogramming, is the step change that has happened in robotics,” says Deepak Pathak, cofounder and CEO of Skild AI, as the company collaborates with NVIDIA to deploy this adaptable intelligence in dynamic environments like factories.

NVIDIA Cosmos and Isaac Tools Accelerate Skild AI Training

NVIDIA Cosmos technologies are central to Skild AI’s ability to translate visual demonstrations into robotic action, diversifying training data and converting video into structured descriptions usable by the S1 model. This process bypasses the traditional need for extensive, manually-collected datasets, a bottleneck Skild AI estimates can require 50 to 100 hours of human effort for a comparable level of training. The collaboration uses Cosmos Curator to annotate, filter, and organize data at scale, accelerating the development of adaptable robotic systems.

This approach is particularly valuable in dynamic environments like manufacturing floors where task variations are frequent and pre-programmed routines quickly become obsolete. Skild AI’s reliance on NVIDIA’s simulation frameworks further streamlines the development pipeline, allowing for rigorous testing and validation before real-world deployment. The NVIDIA Isaac Sim framework provides physically-based virtual environments essential for generating data, identifying potential edge cases, and confirming expected behaviors.

This virtual testing reduces the risk of costly errors and accelerates the learning process, enabling the S1 model to rapidly adapt to new scenarios. Reinforcement learning within NVIDIA Isaac Lab then refines the robot’s capabilities, building upon the foundation established in simulation. Joint development efforts between Skild AI and NVIDIA are focused on new GPU-accelerated simulation solvers designed to accurately model robotic interactions with physical objects.

These solvers, soon to be available as part of Newton, will improve the fidelity of simulated environments, allowing for more realistic training scenarios. Accurate physical modeling is important for tasks involving manipulation, gripping, and the precise handling of materials, ensuring that the robot’s virtual learning translates seamlessly into real-world performance. The integration of NVIDIA Nsight tools and the TensorRT software development kit further optimizes the S1 model for real-world application.

Nsight identifies performance bottlenecks during training, while TensorRT accelerates inference, enabling the robot to respond quickly and efficiently in dynamic environments. This end-to-end optimization, connecting data generation, simulation, training, and deployment, is a key differentiator for Skild AI’s approach, according to the company.

The company’s tests demonstrate a significant performance advantage; in new, multi-step tasks, the S1 robot achieved a 66% success rate at each step, compared to just 9% for a comparable AI system. Skild AI’s ability to achieve functional performance from a single video demonstration is enabled by in-context learning, a technique that allows the model to understand and execute tasks without requiring weight updates or post-training adjustments.

In a plant-potting test, the team transitioned from recording a demonstration to autonomous execution on hardware in just 11 minutes. Skild AI intends to use data from its commercial deployments, where customer agreements allow, to further refine the S1 model, creating a virtuous cycle of continuous learning and improvement. This approach not only accelerates future deployments but also ensures that the robot’s capabilities remain aligned with the evolving needs of its users, solidifying Skild AI’s position in the field of adaptable robotic intelligence.

Skild AI S1 Achieves 66% Success Rate on Novel Tasks

Skild AI’s S1 robot foundation model now achieves a 66% success rate on each step of previously unseen, multi-step tasks, a performance level significantly exceeding comparable artificial intelligence systems which managed only 9% success. The core of S1’s performance lies in a technique where the robot interprets demonstrated intent, objects, and sequences from video input, then maps those observations into appropriate actions without altering its underlying programming, the company says.

Skild AI estimates that a single short video example can provide as much instructional value as approximately 380 manual training examples, a substantial reduction in the time and resources required to deploy robots in dynamic environments.

This efficiency is critical for industries where tasks and layouts frequently change, such as manufacturing and logistics, where traditional robots often require weeks of reprogramming for even minor adjustments. The ability to rapidly adapt to new scenarios allows for greater flexibility and responsiveness on the factory floor.

This capability is not merely about executing known actions in a new order; the robot demonstrates an ability to adjust to unexpected object movements, recover from errors, and combine skills in novel ways. The company’s tests reveal that S1 can successfully navigate tasks involving dozens of manipulation steps, a level of complexity previously reserved for human workers.

With over sixty deployment partnerships established across diverse sectors, including manufacturing, logistics, inspection, security, and even food preparation, Skild AI is rapidly expanding its footprint in a growing number of industries, the company states. The combination of advanced AI algorithms, robust simulation tools, and a commitment to continuous learning positions Skild AI as a key player in the evolving landscape of industrial automation.