A new form of artificial intelligence (AI) enables robots to perform tasks in a much more human way. Typically, robots divide tasks into intermediate goals, just like humans. The AI was developed by researchers at Sweden's Chalmers University of Technology. The AI significantly increases the flexibility of robots and allows them to work closely with humans.
Humans all work in their own unique way; different people solve the same problem in different ways. "An important part of robot development is therefore making robots cooperate with humans in dynamic environments," says Maximilian Diehl, PhD student in electrical engineering at Chalmers University of Technology and principal investigator within the project.
Way we perform tasks varies
The way humans perform simple tasks - think of setting a table - depends on circumstances, among other things. For example, is a chair in the way? Then we push it, move the chair or walk around it. We also use both our left and right hand, take breaks while performing tasks and perform unplanned tasks.
Robots, however, work very differently. They are precisely programmed and carry out concrete instructions. This approach makes them very efficient in environments where fixed patterns can be followed continuously, such as factory environments. However, to interact with people in dynamic environments such as healthcare or a shop, robots need to be able to operate much more flexibly.
"In the future, we foresee robots supporting basic household activities, such as setting or clearing a table, placing dirty dishes in the sink or clearing the groceries," says Karinne Ramirez-Amaro, assistant professor of electrical engineering at Chalmers University of Technology.
Perform tasks in a more human way
In the study, the team therefore looked for a method by which robots learn to perform tasks in a more human way. Here, the robot observes the actions a human performs. The researchers then deploy a so-called 'explainable AI'. This collects generic information about the actions performed. This instead of the specific information that many other AIs collect in this area. The AI uses this generic information to create a flexible and adaptable path to a long-term goal. Explainable AI is a term that refers to a form of AI where people understand how the AI arrived at a particular decision or result.
For the study, several people were asked to perform the same task. This consisted of stacking a series of cubes. Participants performed this task 12 times in a Virtual Reality (VR) environment. Participants' movements were recorded using a series of laser sensors.
Intermediate goals
"If we humans have a task, we divide it into a chain of smaller intermediate goals. And every action we perform is aimed at executing an intermediate goal. Instead of having a robot copy people's behaviour exactly, we focused on identifying the goals, looking at all the actions people performed in the study," says Ramirez-Amaro.
In the method, the AI focuses on mapping intermediate goals. And builds libraries of different actions for each intermediate goal. Then the AI creates a schedule, which can be used by a TIAGo robot. This is a mobile service robot designed for indoor environments. With the help of this tool, the robot could independently create a plan for stacking cubes. Even if the conditions in which this task has to be performed change.
First results are successful
During the trial, the environment in which the robot operates was adjusted slightly each time. The robot could itself continuously choose the right combination of different possibilities actions and place them in the right order to perform the task. The researchers call the results "extremely successful".
"With our AI, the robot could create plans with a 92% success rate after just one human demonstration. If information from all 12 demonstrations, a success rate of 100% was achieved," Diehl said. The next phase of the project focuses on robots' communication with humans. Robots should thus be able to share what went wrong while performing a task and why this is the case.
Applications in industry and healthcare
The researchers hope that robots can be used in industry to perform physically demanding tasks, among other things. Examples include tightening nuts on truck wheels. However, the researchers also foresee applications in healthcare. Including when it comes to distributing and collecting medication and food. "We want to make the work of healthcare professionals easier so they can focus on tasks that require more attention," Ramirez-Amaro explains.
"It may be several years before we actually see autonomous, widely deployable robots. Especially since many individual challenges still need to be addressed, such as computer vision, control and safe interactions with humans. However, we think our approach will help accelerate the learning process of robots. Which enables robots to connect with all these aspects and apply them in new situations," Diehl reports.
Author: Wouter Hoeffnagel
Photo: Gordon Johnson via Pixabay
