Computer scientists from the University of Leeds have used AI techniques to train robots to quickly plan their movements in a messy environment. The team used automated planning and reinforcement learning to train the robots to find an object in a cluttered space, like a fridge or a warehouse shelf, and move the object.

Source: Unsplash

The team wanted to work toward robot autonomy. Robotic autonomy allows machines to access a task and find a solution without human intervention. With autonomy, robots would be able to transfer the skills and knowledge they learned from one task to another.

During the study, the team used a robotic arm confined in an area and challenged the robot to pick up an object on a messy table. To move around the clutter and grasp the goal object, the robot must plan a sequence of moves, including possibly moving other objects out of the way. With current technology, the computer can take several minutes to plan a task, and when the robot does finally decide how it is going to move, it is likely to fail because it doesn't learn from past movements.

Automated planning allows a robot to see a problem through a vision system. The software in the robot’s operating system will simulate a possible sequence of moves it could make to reach an object. But these rehearsed simulations don’t capture the complexity of the real world, which causes the robot to fail a task.

To overcome these issues, the team combined automated planning with reinforcement learning. With this combination of techniques, the computer creates a sequence of possible attempts to reach and move objects. The computer will randomly select a planned move and then use trial and error to learn which planned actions are most likely to be successful. Because it is learning through trial and error, the robot becomes more adept at choosing the most successful option over time.

After applying automated planning and reinforcement learning, the robot could move away from clutter to pick up an object. The team instructed the robot arm to pick up an apple on a cluttered table. The robot quickly analyzed the situation before moving around the clutter to pick up the apple. Decisions that would have taken the robot 50 seconds to analyze before only took five seconds.

This technology will be presented at the International Conference on Intelligent Robotics and Systems.