Researchers have designed an artificially intelligent robot with the ability to learn how to open doors.
For humans, opening a door is a simple, everyday task that we hardly even think about as we go about our lives. For robots, it's far trickier.
Programming a robot to open a door actually involves numerous steps and considerations that are second nature to humans. For example, what is the door type? Does it open inwards or outwards, or slide? If it does open inwards or outwards, then does it swing right or left? Where is the handle positioned, and which is the correct way to turn it?
For computer programmers, opening a door is a difficult task that may need a different approach every time for a robot that is traveling from place to place.
Robots can learn all of these approaches via trial and error, but it might take an awfully long time. One study involved 14 robots grasping an object 800,000 times for two months.
But now, researchers from Japan have taught a robot how to do it by using something called deep predictive learning. Specifically, they designed the robot to be able to split the door-opening problem into multiple parts, called modules, and consider multiple approaches with a predicted success rate.
Robot's Thought Process
The robot's thought process was taught to it by humans who controlled it, allowing its sensors to observe how the task should be performed 108 times with six hours of module training.
At the end of the training, the robot was able to complete its task 96.8 percent of the time. In one test, seen at the top of this article, it went back and forth through the door for around 30 minutes straight.
Not all attempts were successful. The robot tended to fail if the shape and even the color of the door handle was changed from its training sessions. On one occasion the robot grabbed the door handle too lightly and its hand slipped off.
"Although there have been research cases of robot door-opening tasks in the past, they required long development time for modeling the environment and control programming," Tetsuya Ogata, a professor in the Faculty of Science and Engineering at Waseda University in Japan who was involved in the study, told Newsweek. "In contrast, our DPL is based on learning, so the entire task can be accomplished in a very short development time.
"Opening a door is a coordinated movement of the entire body, including the dual-arm arm and the vehicle. One of the most difficult parts is to realize this motion using only the learning method by our DPL. In addition, it was also difficult to achieve robustness that would enable highly reliable execution even in hard situations with changing lighting conditions and human intervention."
In short, the robot's approach to opening doors is far from perfect in the real world. Still, footage of robots opening doors has unsettled some viewers in the past.
In 2018, groundbreaking robotics firm Boston Dynamics released a video of its dog-like SpotMini robot opening a closed door with a jaw-like arm and walking through it. "I, for one, welcome our new robo-dog overlords," one popular YouTube comment read.
Ogata dismissed these concerns. "Our DPL methodology is based on deep learning, which may cause anxiety to users in that its internal mechanisms are difficult to understand," he said.
"Therefore, in this paper, we have made a particular effort to show the results of our analysis and visualization of the internal representation of the learning model. We believe that by making the model as understandable as possible, we can alleviate the above concerns."
The study was published in the journal Science Robotics on April 6.
Uncommon Knowledge
Newsweek is committed to challenging conventional wisdom and finding connections in the search for common ground.
Newsweek is committed to challenging conventional wisdom and finding connections in the search for common ground.
