Engineers from Carnegie Mellon University have unveiled the ALAN robot.AI, Computer Science
Advanced robotic systems available incurrently, are capable of learning, but they need a person to set a task or develop a "reward" system that will motivate research. When creating ALAN, the engineers wanted to develop a system that would explore unfamiliar space on its own, driven only by "curiosity".
The created robot is equipped with a visual module,which can estimate the movements of objects in the surrounding space. Using ready-made, pre-trained detectors, he selects an area to explore, interacts with detected objects, and practices new skills.
Testing the robot's operation in an unfamiliar environment. Video: Carnegie Mellon University
ALAN uses the model of the world formed inlearning process, to determine the actions that can be performed with a particular subject, scientists explain. After that, the robot tests its guesses, constantly expanding its understanding of the surrounding space and its capabilities.
The previously proposed approaches to training autonomous robots require a large amount of initial data, which makes it difficult to develop the technology. The new robot learns continuously without external intervention.
We show that ALAN can learnmanipulate objects with only about 100 trajectories in 1-2 hours in two different game kitchens without any reward. Therefore, the use of visual a priori data can significantly improve the efficiency of robot training.
Russell Mendonka, one of the developers of the robot, in a message to Tech Xplore
Read more:
A treasure with 1,000 coins was accidentally found on a farm: what could they buy
Named a vitamin that protects the brain from dementia
It turned out which men are the most fertile: their sperm is 50% better than the rest
Cover image: Russell Mendonca, Shikhar Bahl, Deepak Pathak