Self-learning robots that predict the results of their actions Wednesday, 06 December 2017

Researchers at University of California Berkeley have developed a robotic learning technology that allows robots to predict the results of their actions, allowing them to manipulate objects they have never encountered before.

The technology, called visual foresight, allows robots to predict what their cameras will see if they perform a particular sequence of movements. It could have future applications in helping self-driving cars anticipate future events on the road, and produce more intelligent robotic assistants.

Visual foresight, in its current form, can only make predictions several seconds into the future, but this is enough to allow the robot to figure out how to move objects around on a table without disturbing obstacles. The robot is able to learn to perform these tasks without any help from humans, or prior knowledge about physics, the environment, or what the objects are. Instead, the visual imagination is learned entirely from scratch, with unattended and unsupervised exploration, where the robot plays with object on a table.

Once the play phase is over, the robot builds a predictive model of the world, and is then able to use this model to manipulate new objects, even those it has not seen before.

“In the same way that we can imagine how our actions will move the objects in our environment, this method can enable a robot to visualise how different behaviours will affect the world around it,” said Sergey Levine, assistant professor in Berkeley’s Department of Electrical Engineering and Computer Sciences. “This can enable intelligent planning of highly flexible skills in complex real-world situations.”

At the core of this system is a deep learning technology called dynamic neural advection, which predicts how pixels in an image will move from one frame to the next based on the actions of the robot. These models have undergone recent improvements, and this, combined with improved planning capabilities, have enabled robotic control based on video prediction to perform increasingly complex tasks, such as sliding toys around obstacles and repositioning multiple objects.

Robots no longer have to rely on human supervisors providing feedback in order to learn skills. Rather, they are able to use the learned model from raw camera observations to teach themselves how to avoid obstacles and push objects around obstructions.

In much the same way that humans learn to manipulate objects through millions of interactions with a variety of objects during their lifetime, the researchers have shown that it is possible to build a robotic system that also leverages large amounts of autonomously collected data to learn widely applicable manipulation skills.