Researchers at Cornwell University Train Robots in Handling Objects

A research team at the Cornwell University led by Ashutosh Saxena, a professor of Computer Science, has observed that of all the tasks that robots can perform, the task of placing objects is probably the mst difficult. This difficulty arises mainly due to the number of choices that are available for the robot for placing the object.

The research team initially experimented with placing plates, glass, bowls, spoons or candy canes on various surfaces like hook, pen holder, dish racks etc. The robot with which they conducted the experiments surveyed the surroundings with the help of a three dimensional camera. After surveying it tried out the various spaces that were available for placing the objects. For some of the cases the robot checked if vertical supports were present to hold the object upright, for some positions it assigned priority according to preferences. After being trained on these lines, the robot was able to place the objects that it had previously seen in their places with 98% accuracy and in case of new objects the robot was correct in 92% of the cases.

For successful placement of objects the researchers have trained the robot to first conduct a survey of the environment before going ahead with the placement of the object. Individual views of the room captured by the camera are brought together to provide a three dimensional image of the room. 24 office scenes and 28 home scenes are fed into the robot with most of the objects labelled. The computer is capable of comparing features like texture, colour etc of the nearby objects with those that are labelled to find similarities. When placed in a new environment the robot compares each of the scanned segments with the previously loaded objects in its memory and tries to match them. The result of these comparisons was that the robot was able to correctly identify 73% of the objects in a home scene and 845 in an office scene. Professor Saxena and his team are planning to give a presentation of some of the findings of the research at the Neural Information Processing Systems conference which will be held during December 2011 at Spain.