A new technology allows robots to predict the outcome of their own actions before they actually perform these actions. It allows robots to independently discover how to manipulate an object, even if they have never seen it before.
It involves a technology called 'visual foresight', developed by researchers at the US University of California Berkeley (UC Berkeley). The technology has been applied to a Baxter collaborative robot from the company Rethink Robotics, which can use it to learn entirely independently without human intervention and without prior knowledge of physics how a particular action affects an object. This enables the robot to manipulate objects independently, even if it is not familiar with the shape, dimensions or other properties of the specific object.
Dynamic neural advection
At the heart of the technology is a deep learning technology called 'dynamic neural advection' (DNA). This technology was developed by UC Berkeley PhD student Chelsea Finn, assistant professor in the UC Berkeley Department of Electrical Engineering and Computer Science Sergey Levine and OpenAI researcher Ian Goodfellow. Finn, Levine and Goodfellow published a research paper on DNA in October 2016. DNA technology divides the image captured by cameras on a robot into different boxes. The model then predicts how pixels of the image will move between boxes if the robot performs certain actions.
In doing so, the robot relies on a predictive model that the robot has generated independently. This model is created based on experiences gained by the robot during a learning phase, which the researchers call the 'play phase'. During this, the robot is presented with a series of objects on a table and instructed to play with them. Cameras record the actions the robot performs with objects and the influence these actions have on the objects. Based on this, DNA technology is used to generate a predictive model. This model then allows the robot to predict how certain actions will turn out before they are actually performed, even on objects the robot has never seen before.
Planning actions intelligently
"In the same way we can imagine how our actions will move objects around us, this method can enable robots to visualise how different behaviours will affect the world around them," explains Levine. "This enables intelligent planning of highly flexible abilities in complex real-world situations." Finn adds, "In the past, robots learned skills from a human supervisor who helped them and gave feedback. What makes this work exciting is that robots can now entirely independently learn a range of skills for visually manipulating objects."
The university points out that the predictions their Baxter robot can make with the DNA model are relatively simple at present. For instance, the robot is able to predict the outcome of its actions only a few seconds in advance. However, this is enough to allow the robot to move objects on a table around a variety of obstacles without colliding with them. In addition, the DNA model allows the robot to reposition multiple objects on a table.
Potential applications
If the technology is developed further, the university foresees interesting applications. For instance, the technology could be used to enable autonomous cars to predict events on the road in advance. DNA models could also help intelligent robotic assistants operate smarter in home situations.
"Children can learn about their world by playing with toys, moving them back and forth, grasping them, and so on. Our goal with this research is to enable a robot to do the same thing: to learn how the world works through autonomous interaction," Levine said. "This robot's capabilities are still limited, but its skills are learned entirely independently and enable it to predict complex physical interactions with objects it has never seen before by building on existing observed patterns of interactions."
"Humans learn skills to manipulate objects without a teacher using millions of interactions with diverse objects they have throughout their lives. We have shown that it is possible to develop a robotic system that also uses a large amount of autonomously collected data to learn widely applicable manipulation skills, specifically skills to push objects," said Frederik Ebert, a UC Berkeley PhD student.
Model further development
Researchers at UC Berkeley want to further develop the model and build more sophisticated models for robots. These models should start enabling robots to perform more complex tasks, such as picking up and putting down objects, manipulating soft or deformable objects such as clothing or rope, and assembling products.
Author: Wouter Hoeffnagel
Source: University of California Berkeley
Source photo: Pixabay / DirtyOpi
