Alphabet, Google's parent company, is experimenting with robots capable of learning tasks on their own without having to be programmed to do so. According to the company, this is necessary to achieve robots that can effectively support people in their daily lives.
The project is called Everyday Robot Project and was announced by Hans Peter Brondmo, general manager of the project. "Today's world of robots is very similar to the world of computers of 50 years ago. It is a hot topic and there is a lot of optimism about how robots can help in people's daily lives. However, this future is still far away. Most robots are at the stage where mainframe computers were in the 60s and 70s. They are expensive specialised machines managed by experts and perform specialised tasks is specially designed environments," Brondmo writes in a blog post.
Combining vision, comprehension, navigation and actions
It points out that humans naturally combine sight, understanding, navigation and actions to achieve goals. Contemporary robots cannot do this and have to be accurately programmed. This makes it very challenging for robots to perform tasks that are easy for humans in a highly dynamic and changeable environment.
"To make robots useful in everyday environments, we need to move away from precisely coding robots to perform specific and structured tasks in exactly the right way and at exactly the right time," Brondmo said. Machines must learn to perform sensible tasks on their own, he argues.
Active behind the scenes for years
Although the Everyday Robot Project has now been officially announced, the team has been working behind the scenes for several years. Several of the team’s experiments have, however, been made public. For example, the team previously demonstrated how robots can learn more quickly when simulated in the cloud, how robots can learn from human demonstrations, and how robots can share experiences to learn from one another.
The knowledge gained by the team with this is now being put to practical use. The team teaches robots to sort waste, distributing cups, bottles, packaging and other materials to landfills, recycling or compost heaps. This uses various machine learning techniques, including simulation, reinforcement training and collaborative training.
Training in the cloud
The robots are trained in the cloud; every night, tens of thousands of virtual robots train in a virtual office to further refine their sorting skills. This cloud-based training is combined with on-the-job training using physical robots. The data gained in this process is fed back to the cloud and the virtualised robots, which take this input into their virtual training. As a result, the robots learn from each other.
The results are promising, according to Brondmo. "In recent months, our robots have sorted thousands of pieces of waste and reduced the pollution level (ed: the amount of mis-sorted waste) from 20% - which is what it is when people throw objects in the rubbish - to less than 5%," he writes.
"The results are promising for several reasons. Firstly, they show that we are capable of creating a robotic system that integrates all the robot's functionalities to do something actually useful: reducing pollution levels to less than 5% successfully keeps waste away from landfill. In addition, it proves that robots are capable of learning new tasks in the real world through practice, rather than engineers 'handing over code' for every new task, exception or improvement."
Learning other tasks
In the coming time, the team wants to try to apply the knowledge gained by the robot in performing this task to teach another task, without rebuilding the robot or writing a lot of new code. "This may remain impossible, but we are going to try," Brondmo said.
The full blog post can be found here.
Author: Wouter Hoeffnagel
