Babies spend their first few months learning how to find their way around their environment and how to use objects. They are also very talented at this. For example, babies who learn that mug glasses are in different shapes and sizes, but they all have to hold, decide that they can hold jugs by analogy.
The team led by Ashutosh Saxena, who works as a computer science specialist at the Cornell University Personal Robotics Lab, teaches robots to use objects and find their way in new environments.
The main theme of the research presented in 2011 Robotics: Science and Systems Conference is to produce learning machines using computers that plan to observe events and find similarities. For example, with the right programming, it is possible to have the computer examine a wide range of mugs, find the common properties of the mugs and enable it to recognize other trophies in the future. With a similar process, it is possible to have a robot find the handle of the cup and hold it properly.
Other researchers have reached these conclusions before, but unlike previous teams, Saxena's team focused on placing objects rather than holding them. Placing objects is harder than holding them, because there are more options for placing objects. For example, we put a mug flat on the table, but if we wanted to put it in the dishwasher, we would do it upside down. For a machine that can learn, this corresponds to a new learning process.
The innovation brought by the team's work to the world of science is to teach a robot contextual relationships in 3 dimensions. The robot that successfully identifies objects at 83 percent in home tests and 88 percent in office environments. In his last test, he was able to identify the keyboard in a room he never knew.
In the first tests, robots were able to correctly place objects such as plates, mugs, martini glasses, bowls, candy canes, discs, spoons and tuning fork on many dish racks, such as the flat surface in the dishwasher, goblet holder, hanging hook, cutlery. Robots do this by observing their surroundings with a 3D camera and random testing for proper placement. He is investigating the presence of supports for placing objects and prefers the best placement in a priority order. For example, while placing the plate on the table, he prefers to place it horizontally, while he prefers to place the plate upright in the dishwasher.
The secret: 3D contextual relationship
Just as we unwittingly identify objects when we enter a room, the robot developed by Saxena and his team likewise scans the room it enters and identifies the objects in it. It creates a 3-dimensional view of the whole room by combining the images taken with its 3D camera. He then divides the image of the entire room into sections, evaluating the discontinuities and distance between objects. The purpose of the robot to perform these operations is to label each part separately.
The researchers were able to train a robot by giving it 24 office and 28 home environments where most objects were tagged. The computer managing the robot decides what the common properties are by examining the characteristics of objects with the same label, such as color, texture, and what's close. In a new environment, he compares the sections formed from the 3D view of the environment and the objects he distinguishes within the sections with the most appropriate objects he identifies in his memory.
The innovation brought by the team's work to the world of science is to teach a robot contextual relationships in 3 dimensions. The robot that successfully identifies objects at 83 percent in home tests and 88 percent in office environments. In his last test, he was able to identify the keyboard in a room he never knew. Contexts give their robots an advantage, Saxene says, “The keyboard appears in just a few pixels in the images, but the monitor is easily found. To find a keyboard, it may be easier to find the monitor first. Because keyboards are located under the monitors. Here is my robot was able to locate the keyboard using this information. " he explained.
Still, Saxena's team admitted that despite all the progress in their work, they had a long way to go. "I'd be really happy if they could build a robot that could act like a 16-month-old baby," said Saxena.
After the robots developed in Cornell Laboratories are trained, they place the objects they have seen before with 98 percent accuracy in the environment where they are trained. These robots can place new objects they encounter in a new environment with 95 percent accuracy. Researchers say that with longer training, performance will increase even more.