Category: News

3.Versatile Robots on the Horizon: Embracing the Multitasking Era

Drawing inspiration from the foundational principles fueling the current surge in Generative AI, roboticists are embarking on the creation of more versatile robots capable of handling a diverse array of tasks.

In recent years, the landscape of AI has shifted from employing multiple specialised models, each dedicated to distinct tasks such as image identification, drawing or captioning. Instead, there’s a move towards monolithic models, exemplified by OpenAI’s GPT-3, capable of being fine-tuned to tackle coding problems, scriptwriting, high school biology exams and beyond. 

Multimodal models like GPT-4 and Google DeepMind’s Gemini demonstrate proficiency in both visual and linguistic tasks. This approach is now extending to robotics, eliminating the need to train separate robots for distinct functions like flipping pancakes or opening doors. A one-size-fits-all model could empower robots to multitask efficiently, as evidenced by several noteworthy developments in 2023.

In June, DeepMind introduced Robocat, an evolution of the previous year’s Gato, designed to autonomously generate data through trial and error, enabling it to control a variety of robot arms instead of a specific one.

Come October, DeepMind unveiled another general-purpose model for robots, RT-X, along with a substantial collaborative effort involving 33 university labs to create a comprehensive general-purpose training dataset. Other leading research teams, including RAIL at the University of California in Berkeley, are exploring similar technologies.

However, a key challenge lies in the scarcity of data. While Generative AI benefits from vast internet-scale datasets of text and images, robots face a dearth of reliable sources to learn industrial or domestic tasks. Addressing this issue, Lerrel Pinto at New York University and his team are pioneering techniques that allow robots to learn through trial and error, generating their training data in the process.

This innovative approach, already proving successful in the realm of driverless cars, is exemplified by startups such as Wayve, Waabo, and Ghost. They are leading a new era of self-driving AI, using a single large model to control a vehicle rather than multiple smaller models for specific driving tasks. This shift empowers smaller companies to compete with industry giants, akin to Wayve testing its driverless cars on the bustling streets of London.

As the paradigm of large-scale models revolutionises the field of driverless cars, a similar transformative boost is expected to ripple across robotics, ushering in an era of versatile and efficient multitasking robots.

“The synergy of generative AI principles with robotics is propelling us into a new era of versatile automation. The concept of employing a single model for multitasking robots is a game-changer, ushering in a dynamic future for the field,” notes Mark Kelly, Founder of AI Ireland.

Do you agree? Come back next year and check how we did.