Berkeley researchers develop robot that teaches itself tasks




Researchers at the University of California – Berkeley have developed a robot who learns how to complete certain tasks. That is thanks to a new “deep-learning algorithm, which uses simulated neural networks.

The researchers wrote two papers on the project , which they next Wednesday and Thursday will present at the International Conference on Robotics and Automation in Seattle. The advantage of a robot that uses “deep learning” is that he does not need to be reprogrammed for each task. Instead, the software allows the robot tasks can learn. That way of functioning to make robots suitable for use in unpredictable environments. Now robots are mainly used in controlled environments in which objects are always in the same place, because they can not automatically adapt to changes in the environment.

The researchers worked with a Personal Robot 2 , nicknamed Brett or Berkeley Robot for the Elimination of Tedious Tasks. The PR2 was developed by Willow Garage, a California robot builder. The 133 to 165cm long pull-out apparatus consists of a torso with two gripping arms. In the ‘head’ are include Microsoft Kinect and a laser scanner. The ‘brain’ of the device is formed by two quad-core Intel Xeon processors with 24GB memory.

The deep-learning algorithm that the researchers developed, uses convolutional neural networks with 92,000 parameters. Such algorithms make use of simulated neurons, each of which process a small piece of the raw data. Because the neurons partly overlap each other, they are also able to recognize larger patterns in the data. Convolutional neural networks are already being used in other speech and image recognition.

The algorithm uses camera images and the position of the pivot points of the arms to provide real-time feedback on the progress of the task in the form of a score. Movements that bring the robot closer to the execution of the task, thereby score higher than other movements. In this way, the robot will learn the best way to carry out a task.

The researchers Brett left to perform a number of motor tasks, such as screwing a cap on a bottle or the stacking of LEGO blocks on each other. By the algorithm, the robot could learn such tasks in about ten minutes, when the researchers gave him advance the coordinates of the various objects in the environment. When Brett had to recognize the objects in the scene itself, it took him about three hours to learn a task.

The researchers are optimistic about the potential of a robot that learns with deep learning tasks. They expect that such techniques robots on a period of five to ten years, set in a position to carry out complex tasks.


In: A Technology & Gadgets Asked By: [19122 Red Star Level]

Answer this Question

You must be Logged In to post an Answer.

Not a member yet? Sign Up Now »