Wouldn’t it be nice if a robot could learn how to use an object without any help from humans? Apparently, that’s what a robot called Vestri can do.
Developed by a team from the University of California in Berkeley, Vestri can “see” into the future – but just for a few seconds – with that foresight allowing them to maneuver objects they have never been in contact with. The technology making it possible is called “visual foresight”, which may one day be used on self-driving cars or used to create home assistants that are more intelligent.
The researchers from UC Berkeley took inspiration from how children play. Kids play around with toys to figure out how it works. That very concept was programmed into the robot. So once the robot is done playing with an object, it then creates a predictive model which it taps into to perform a task.
The robot is helped by cameras, which help it see the next sequence of events. It uses the cameras to create different scenarios that haven’t happened yet. It “imagines” the possible events then picks the most effective method to achieve something. For example, moving an object from one place to the next.
Sergey Levine, an assistant professor at Berkeley, said in a university press release that the method allows the robot to visualize how different behaviors will affect the surrounding environment. This ability allows for “intelligent planning of highly flexible skills in complex real-world situations.”
The methods used are different from conventional computer-vision models, where lots images numbering in the thousands or millions are labeled and programmed into the machine. Instead, the method used for Vestri involves information that is autonomously collected.
The technology developed by researchers at UC Berkeley can be used on self-driving cars in the future. A car with such a technology can anticipate events that may happen on the road. But that is a plan for the future. For now, researchers are focusing on how Vestri can learn simple manual skills from autonomous play.
What Vestri can do right now is still very simple, but it’s sufficient enough to allow them to move objects around without any obstacles.
A deep learning technology lies at the heart of Vestri’s system, and it is based on convolutional recurrent video prediction, or dynamic neural advection (DNA). DNA-based models are able to predict how image pixels will move from frame to frame based on the actions of the robot. Improvements to DNA-based models have allowed robotic control that is based on video prediction to perform complex tasks such as sliding toys around obstacles.
Before this, robots learned skills with the help of humans who provided feedback. This is why the work done with Vestri is so exciting as it has shown that robots can learn without needing human assistance.
UC Berkeley researchers will continue to study control via video prediction, but further research will involve on improving prediction and prediction-based control, as well as coming up better methods that allow the robot to collect more focused video data so they can perform even more complex tasks like picking and placing objects, soft object manipulation, and assembly.