Skip to main content

MIT researchers teach a robot to teach other robots

Get updates direct: Subscribe via email
Tinuku ~ Researchers at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) developed a system called C-LEARN to allow noncoders to teach a robot about various tasks just by giving some information and other robot demos are performing the task.

Tinuku MIT researchers teach a robot to teach other robots

MIT's CSAIL developed a system allows users to teach skills to a robot to be automatically transferred to other robots that have different ways of sizing time and save costs for companies who want a variety of robots to perform similar actions in the real world.

"The combination of learning intuitions uses demonstrations with precision motion planning algorithms," says Claudia Pérez-D'Arpino, PhD student at MIT.

"Robotic approaches to perform new types of tasks they have never learned before, such as multistep assembly using both arms," Pérez-D'Arpino, says.

The research team tested the system on Optimus is a robot designed to isolate bombs and perform tasks such as opening doors, hauling objects, and removing objects from containers. Optimus's learning abilities transferred to the Atlas is a 6-foot high humanoid robot and weighs 400-pounds.



C-LEARN provides a knowledge base of information on how to reach and understand different objects that have different constraints. For example, tires and steering wheel have a similar shape, but to attach it to the car, the robot must adjust the arm differently.

The operator then uses a 3-D interface for a demonstration robot about a specific task indicated by a sequence of relevant moments known as "keyframes" for different situations. The robot automatically suggests a motion plan for the operator to approve or edit as necessary.

"This approach is actually very similar to humans learning something about the world, we can not miraculously learn from a demonstration, so we take new information and match it with previous knowledge of the situation and environment," says Pérez-D'Arpino said.

This system is part of a focus on adaptive learning approach. C-LEARN handles certain advanced tasks such as avoiding collisions or planning different sequence of steps for a specific task. But the team hopes for more insight that gives the robot a wider physical ability.

"Traditional programming of robots in real-world scenarios is difficult, tedious, and requires a lot of domain knowledge," said Julie Shah, MIT Professor.

"It is far more effective to train them by providing basic knowledge and a single demonstration.This is an exciting step towards teaching robots to perform the multiarm and multistep tasks required for the assembly and maintenance of ships or aircraft," Shah said.

Comments

Popular posts

Gesits electric motorcycle setup 100,000 units in 2018

Walmart, Nestle and Unilever join IBM foodchain

Ticket Camp announced the use of bitcoin

China established first cyber-court in Hangzhou

SK Planet and Salim Group rumored in talks on Elevenia acquisition