By Chris Skeldon 
All about artificial intelligence
The Toyota Research Institute (TRI) has revolutionized the way robots learn to perform complex tasks, such as preparing breakfast, through the use of artificial intelligence (AI), as seen in a video released by the company to demonstrate the innovative training technology. .
In an environment they describe as “a robotics kindergarten,” TRI researchers are implementing technologies that promise to eliminate the need for hundreds of hours of complex programming and troubleshooting.
Read more:
The key to the success of this process is providing the robots with a sense of touch. Equipped with a kind of “soft thumb,” robots can “feel” what you are doing, providing important information to carry out difficult tasks that would otherwise be difficult to perform based on vision alone.
It’s exciting to see them interact with their surroundings.
Ben Burchville, Director of Handling at Toyota Research Institute
The process begins with a human “teacher” demonstrating a series of skills to the robots. Then, over the course of a few hours, the AI model learns these skills in the background. “It’s common for us to teach a robot in the afternoon, let it learn overnight, and the next morning we’re faced with a new functional behavior,” Burchville adds.
Researchers are committed to creating “large behavior models” or LBCs (large behavior models) for robots. These models, similar to large language models, also called LLMs (large language models, in English) used to create human text, will learn through observation and be able to perform new tasks that have not been explicitly taught before.
This approach is revolutionary for robotics, explains Ross Tedrick, a professor of robotics at MIT and vice president of robotics research at TRI. Using this process, the researchers claim to have successfully trained more than 60 difficult skills, including “pouring liquids, using tools, and manipulating deformable objects.” They aim to increase this number to 1,000 by the end of 2024.
It is worth noting that other technology companies, such as Google and Tesla, are also exploring similar approaches. Just like the Toyota researchers, their robots use the experience they gain to infer how to perform tasks.
In theory, AI-trained robots could, in the future, perform tasks with little or no instruction, similar to the direction that would be given to a human (“clean up that leak” for example).
main achievements
The Toyota Research Institute describes the greatest achievements of its work to develop grand models of behavior as follows:
- Publication policy: TRI and collaborators in Professor Song’s group at Columbia University have developed a powerful new approach to behavior learning based on generative AI called “diffusion policy” that enables fast and easy teaching of behaviors through demonstrations.
- Custom robot platform: TRI’s robot platform is specifically designed for flexible two-arm manipulation tasks, with a particular focus on haptic feedback and haptic sensing capabilities.
- Pipeline: TRI robots have already learned 60 flexible skills, with a goal of reaching hundreds by the end of the year and 1,000 by the end of 2024.
- Drake: Part of the Toyota Research Institute’s “not-so-secret” is Drake, a model-based robotics design that provides an advanced tool and simulation platform. Drake’s high accuracy allows you to develop in simulation and in reality at a significantly increased scale and speed. The in-house robotics group was built using Drake’s systems and optimization frameworks. TRI is making Drake open source to stimulate action across the robotics community.
- protection: Safety is paramount in TRI’s robotics efforts. The institute designed the system with strong safeguards, backed by Drake and the robot’s custom control group, to ensure the robots respect safety guarantees such as avoiding collisions with themselves or the environment.
We have a lot of work ahead of us
- However, as he pointed out New York times When dealing with Google search, this type of work tends to be “slow and tedious.”
- Providing sufficient training data is much more difficult than simply feeding an AI model with a large amount of data from the Internet.
- The New York Times article highlights an example where a robot mistakenly identified the color of a banana as white.
- This shows the challenges they still face in developing these advanced AI systems.
- However, advances in robots’ ability to learn through experience and observation promise to revolutionize the automation of complex tasks, making it easier and more effective than ever before.
- Research by Toyota and other companies in this field continues to shape the future of robotics and artificial intelligence.
Have you watched the new videos on Youtube From Olhar Digital? Subscribe in the channel!