In the future, this technology could help self-driving cars anticipate future events on the road and produce more intelligent robotic assistants in homes, but the initial prototype focuses on learning simple manual skills entirely from autonomous play.
Using this technology, called visual foresight, the robots can predict what their cameras will see if they perform a particular sequence of movements. These robotic imaginations are still relatively simple for now - predictions made only several seconds into the future - but they are enough for the robot to figure out how to move objects around on a table without disturbing obstacles.
Crucially, the robot can learn to perform these tasks without any help from humans or prior knowledge about physics, its environment or what the objects are. That's because the visual imagination is learned entirely from scratch from unattended and unsupervised exploration, where the robot plays with objects on a table. After this play phase, the robot builds a predictive model of the world, and can use this model to manipulate new objects that it has not seen before.
The research team will perform a demonstration of the visual foresight technology at the Neural Information Processing Systems conference in Long Beach, California, on December 5.
At the core of this system is a deep learning technology based on convolutional recurrent video prediction, or dynamic neural advection (DNA). DNA-based models predict how pixels in an image will move from one frame to the next based on the robot's actions.
Recent improvements to this class of models, as well as greatly 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.
With the new technology, a robot pushes objects on a table, then uses the learned prediction model to choose motions that will move an object to a desired location. Robot use the learned model from raw camera observations to teach themselves how to avoid obstacles and push objects around obstructions.
Since control through video prediction relies only on observations that can be collected autonomously by the robot, such as through camera images, the resulting method is general and broadly applicable. In contrast to conventional computer vision methods, which require humans to manually label thousands or even millions of images, building video prediction models only requires unannotated video, which can be collected by the robot entirely autonomously. Indeed, video prediction models have also been applied to datasets that represent everything from human activities to driving, with compelling results.
The Berkeley scientists are continuing to research control through video prediction, focusing on further improving video prediction and prediction-based control, as well as developing more sophisticated methods by which robots can collected more focused video data, for complex tasks such as picking and placing objects and manipulating soft and deformable objects such as cloth or rope, and assembly.