Learning from Mistakes: Worker Robots

2020-01-31

"Persistence pays off" - this is a saying that helps humans become highly skilled, and now this method is being applied to robots. Computer scientists at the University of Leeds are using artificial intelligence (AI) technology for automated planning and reinforcement learning to train robots to find and manipulate objects in cluttered spaces such as warehouse shelves or refrigerators. The goal is to develop autonomy in robots so that they can assess the unique environments presented in tasks and find solutions, similar to how robots transfer skills and knowledge to new problems. The biggest challenge is that in a confined area, a robot arm may not be able to grasp an object from above. Instead, it must plan a series of movements to reach the target object, possibly by manipulating other items. The computational power required to plan such a task is so great that robots often pause for several minutes. When they actually perform the action, they often fail. The computer scientists at the University of Leeds combine two ideas of artificial intelligence to develop the concept of "practice makes perfect". The first is automated planning. The robot is able to "see" the problem through a visual system, essentially an image. The software in the robot's operating system simulates the sequence of actions it may take to reach the target object. However, the simulated rehearsals by the robot fail to capture the complexity of the real world, and when they are implemented, the robot is unable to perform the task. Therefore, the Leeds team combines planning with another artificial intelligence technique called reinforcement learning. Reinforcement learning involves a series of trial and error attempts by the computer - about 10,000 in total - to reach and manipulate objects. Through these trial and error attempts, the robot "learns" which actions in its plan are more likely to succeed. The computer learns on its own, starting from randomly selecting a potentially effective plan of action. However, as the robot learns from trial and error, it becomes better at selecting plan actions with a higher chance of success. Dr. Matteo Leonetti from the School of Computing said, "Artificial intelligence excels at making robots reason - for example, we have seen robots participate in international chess matches with grandmasters. But robots are not good at tasks that humans excel at: high mobility and dexterity. These physical skills are hardwired into the human brain, a result of evolution and our way of practicing, practicing, and practicing." "This is an idea we are applying to the next generation of robots." According to Wissam Bejjani, the doctoral student who wrote this research paper, the robot develops a generalization capability to apply what it plans to a set of unique environments. He said, "Our work is significant because it combines planning and reinforcement learning. Many research efforts attempting to develop this technology have focused on one of these methods. Our approach has been validated by the results we have seen in the university's robotics lab." Dr. Mehmet Dogar, an associate professor in the School of Computing, also contributed to this research. He said that this approach speeds up the "thinking" time of the robot by 10 times - a decision that used to take 50 seconds now only takes 5 seconds. This research is funded by the UK Engineering and Physical Sciences Research Council, as part of a project to study "human-like physics" in robotics.

RECENT AI NEWS

RECENT AI TOOLS