Skild AI’s Skillful and Scalable AI Foundation Model for Robotics
Robots require brains, whether they are four-legged military “robot dogs” meant for urban fighting or bipedal humanoids performing routine industrial jobs. These have always been quite specialized and designed with a specific function. However, a robotics startup based in Pittsburgh, SKILD AI, asserts that it has developed a single off-the-shelf intelligence that can be installed into many robots to enable fundamental features.
Skild AI can be installed into a range of robots to do tasks like navigating over obstacles in their path, climbing steep slopes, and identifying and picking up objects. The basic model of Skild AI, trained on 1000 times more data than its competitors, claims its founders.
Abhinav Gupta and Deepak Pathak, two former professors at Carnegie Mellon University, founded this Skild AI in May 2023. The company has developed a foundational model for what it calls a “general purpose brain” that can be installed in a range of robots to perform tasks like climbing steep hills, navigating obstacles in their path, and recognizing and picking up objects.
Many prestigious venture capital firms, including Lightspeed Venture Partners, Coatue, Softbank Group Corp., and Jeff Bezos’s Bezos Expeditions, led the first series of investments. Participating organizations included: Carnegie Mellon University, Felicis Ventures, Sequoia, Menlo Ventures, General Catalyst, CRV, Amazon, and SV Angel.
From Data to Dexterity: Large-Scale Models Powering Advanced Robotic Learning
Deepak Pathak, the CEO and co-founder of Skild AI, said that the large-scale model they are building demonstrates exceptional flexibility and impromptu skills for a variety of robots and jobs. This methodology can significantly improve automation in real-world scenarios. According to Deepak, Skild AI represents a significant breakthrough in robotics scalability and has the potential to completely transform the physical economy.
By training its model on a big database of text, photos, and video, Skild has been able to accomplish this feat. The company says that this database is multifold than its competitors utilize. In order to generate this enormous database, the cofounders, who were both formerly employed as Artificial Intelligence researchers at Meta, combined a variety of data collection methods that they had created and tested over the course of several years of research.
They hired contractors to operate robots remotely and collect data for innumerable tasks, and engaging the robot through a series of random tasks, recording the outcomes, and learning through trial and error. In addition, the AI model was trained using footage from millions of public recordings.
Deepak came up with a method to teach robots “artificial curiosity” while he was a PhD student at UC Berkeley. He did this by giving them rewards for achieving accomplishments even when they couldn’t have predicted them. “An agent’s curiosity increases in relation to the degree to which it is unsure of the predicted effect of its actions,” he said. Because of this method, the AI was able generate more data.
More than 4,000 people have referenced his 2017 study on curiosity-driven learning, he revealed. Robots can now take written instructions from huge language models like GPT and turn them into actions, thanks to Deepak’s innovation. He emphasized, “We found a way to integrate all of these components into a unified system in 2022, which is the idea of combining knowledge from simulation with learning from videos, from curiosity, and real data.”
Similar to OpenAI, Deepak and Abhinav see their company’s future in which Skild’s core model may be fine-tuned to support a variety of use cases and products. Cofounder Abhinav stated, “This is precisely how we hope to transform the robotics sector.” He also mentioned that their ultimate goal is to develop a theoretical artificial general intelligence system, that can equal or exceed human robotics capabilities and that people will be able to interact with in the real world.
He asserts that, Skild AI stands apart from competitors due to its access to vast amounts of data, and that the secret to scaling robotics is data. A particular kind of data that isn’t readily available online is needed by robots. It’s not always possible to apply simulation data to real-world situations. Robotics is currently quite enthused about the prospect that something like large language models and large vision language models can be archived, with billions of examples of internet-scale data accessible for both.
Sustainable Robotics: Skild AI and the future of creativity
This future is closer than ever, thanks to AI. By infusing robots with artificial intelligence, we’re on the verge of immense technological advancements in this 21st century, AI empowering robots with capabilities like machine learning, and allowing them to learn from experience and improve over time.
With the advent of SKILD AI and the like, the synergy between AI and robotics will go beyond efficiency. AI can equip robots with social intelligence, enabling them to interact and collaborate with humans more effectively. This paves the way for robots to assist tackling complex tasks in dynamic environments, too.
The future of AI in robotics is brimming with potential. From boosting human productivity to enabling groundbreaking scientific discoveries, this powerful combination holds the key to a brighter future for all.
ALSO READ:
Will OpenAI Sora text-to-video model Revolutionize Video Creation in 21st century
Alaya AI: Revolutionizing the incredible Future of Artificial Intelligence
SUNO AI: The Powerful impact on artificial Music Generation
