Robot learns to open doors by splitting the task into three easy steps

Splitting a robot’s AI brain into modules that handle simpler tasks means it can be trained more quickly, but that may come at the cost of adaptability

By Matthew Sparkes

A robot has learned to open doors using a new method that reduces the time and effort required to train it, but that efficiency may come at the cost of adaptability.

Robots are often controlled by a deep learning model that has been trained over thousands of trial-and-error attempts to complete the task. Instead, Hiroshi Ito at Waseda University, Tokyo, and his colleagues split the model into modules, with one controlling the robot as it approached the door, another taking over to open the door and one handling passing through the entrance. For each task, the robot had one module for doors that pull open and one for doors that push open.

The robot received 6 hours of training for each of the six modules and was shown how to perform the task by humans 108 times. This is less training overall than a single model would need because each module was trained on a smaller, simpler task. Ito says that a comparable problem by Google researchers took two months of training, using 14 robots in parallel.

After training, the robot accomplished its task 96 per cent of the time. In one test it went back and forth through the door for 30 minutes straight, completing 15 round trips.

The robot runs all of its modules continuously. Each one suggests what it should do next, and an “operation selector” chooses the most appropriate action for the situation and switches from one module to another as appropriate. The team suggests that this could increase adaptability, because rather than training an entire model to work with a new type of door, a module to open that door could be slotted in.

Sethu Vijayakumar at the University of Edinburgh, UK, says the approach has merit, but one large model can learn additional tricks that may improve its performance, whereas separate modules are limited in what they can learn. For instance, a single model could observe important details about the door handle as it approached it, whereas a single module that attempts to open the door once the robot has arrived wouldn’t see these details.

“I believe that this would have improved the data efficiency of the method. What I’m still very sceptical about is the generalisability,” he says. “There is no such thing as a free lunch in machine learning.”

Journal reference: Science Robotics, DOI: 10.1126/scirobotics.aax8177