Purpose-built robots, designed to perform only specific tasks, are facing extinction as a general-purpose brain for androids is on the way courtesy of Skild AI, which has secured $300 million in funding.
The Pittsburgh-based robotics amateur announced that it has created a foundational model for its “general purpose brain” that can be fitted into various robots. According to the company’s press release, these machines range from “quadrupeds mastering adverse physical conditions” to “vision-based humanoids performing dexterous manipulation of objects for complex household and industrial tasks.”
The single off-the-shell intelligence allows robots to accomplish various basic functions, such as maneuvering a path with obstructions, picking up objects, and climbing.
Building the Brain
Although quite new to the robotics arena since its start in May 2023, Skild AI trained its robot brain system using a massive database of text, images, and video. The company claims this database is 1000 times larger than its rivals.
Co-founders Abhinav Gupta and Deepak Pathak, former professors and researchers at Meta, drew inspiration from their years of experience and built the database by combining multiple data-gathering techniques.
One of the company’s machine-learning approaches included using human operators who controlled the robots remotely so the AI could learn from a real-life person. Another was through trial and error, wherein the robots were assigned random duties and allowed to learn through multiple attempts.
Pathak also incorporated his research on curiosity-driven learning into the development, teaching robots “artificial curiosity” to motivate them to explore more outcomes and collect more data. “The more uncertain the agent is about the prediction of the effect of its actions, the more curious it gets to explore,” he explained.
On top of that, robots could interpret texts from large language models, such as ChatGPT, and convert them into actions, according to Pathak. “In 2022, we figured out a way to put these things together in a single coherent system,” Pathak said. “The notion of learning from videos, learning from curiosity, learning from real data but combined with the knowledge from simulation.”
Emergent Behavior
The AI training seemed to have paid off as Skild’s general-purpose brain enabled robots to have an emergent behavior, demonstrating actions that were not part of their training.
Among these actions were rotating an object and recovering an object that slipped. These behaviors, while nuanced, are similar to how humans perform subconsciously using subtle motor skills, which promises flexibility.
The same emergent behavior had been observed from the humanoid robot of fellow startup Figure AI, but this droid is a custom-coded machine that only operates using an external AI provider.
Backed by Big Names
The impressive feat that comes with building a one-size-fits-all robot brain did not fail to catch the attention of investors, as Skild AI received $300 million in a Series A funding dominated by venture capital names like Lightspeed Ventures, Japan’s Softbank investment company, and Amazon’s Jeff Bezos.
Felicis Ventures, Menlo Ventures, CRZ, Sequoia Capital, General Catalyst, SV Angel, and CMU also put their hands in the business, turning the novice Skild AI into a robotic powerhouse with a $1.5 billion valuation.
