Skild AI is developing the 'Omni Brain,' a universal, general-purpose AI brain designed to power any robot across any form factor for any task. They address the fundamental 'data problem' in robotics by leveraging a diverse range of data sources—robot, video, and simulation—through a pre-training and post-training methodology, aiming to create a self-sustaining 'data flywheel' for rapid deployment and continuous improvement.
Summarized by Podsumo
Robotics faces a unique **'data problem'** due to the lack of a vast 'internet of robot data,' unlike language or vision, necessitating a generalist approach to data utilization.
Skild AI's **Omni Brain** is a 'one brain, any robot, any task' solution, aiming to be the horizontal platform for robotics, much like LLMs are for language, by making every deployment contribute to the brain's overall intelligence.
The company employs a multi-faceted data strategy, using **videos for pre-training** (general understanding), **simulation for practice and robustification**, and **real-world robot data for post-training** (precision and deployment-specific fine-tuning).
A key concept is the **'data flywheel,'** where data from specialized robot deployments across various sectors (factories, hospitals, homes) feeds back into the general brain, accelerating learning and enabling automation of increasingly complex tasks.
Skild AI partners with NVIDIA, utilizing technologies like **Isaac Sim for physics simulation**, generative AI models for data augmentation, and NVIDIA's compute platforms, especially for **on-device edge compute** critical for real-time robot reactions.
"Robotics is a data problem. Unlike language or vision, there is not much data in robotics. There is no internet of robot data."
"Think of like what chat GPT is for language. We are building a general brain for any physical device or any kind of robot."
"If we can learn everything from videos, Deepa give this is gives us this a great example that if we can learn from videos all of us would be feders."