Just as a toddler will quickly learn to recognize a tasty treat or avoid grasping a prickly object, software released today will make it easier to teach robots to learn how to behave through experience rather than painstaking reprogramming.
Such software could ultimately make robots simpler to use. Instead of writing new code or accessing a graphical user interface, with BrainOS a user could train a robot to perform a simple task, such as moving toward a particular object, by guiding it through the process manually first.
It is especially difficult for robots to perform reliably in complex, changing settings, which is why learning is a promising approach. Some commercial robots are already capable of simple feats of learning, in fact. But BrainOS includes a range of techniques for learning, making it easy for robot engineers to tap into software libraries for learning tasks such as object recognition, navigation, and manipulation.
Brain Corporation demonstrated the software at a robotics event in Boston last year, using a small red prototype robot resembling a Segway with two cameras for its eyes. It was possible to show the robot an object, and then have it follow that object around.
The learning capabilities used by BrainOS involve feeding information into a simulated network of virtual neurons and synapses, and then providing positive or negative feedback, a process known as “supervised learning.” This approach has proved especially effective in recent years.
Todd Hylton, executive vice president of Brain Corporation, said in a statement that machine-learning software is mostly aimed at academic researchers rather than engineers in industry. “BrainOS helps solve that problem by providing a central technology framework that is accessible to those looking to prototype, commercialize, and monetize robotic applications,” Hylton said.
Ashutosh Saxena, director of a project called RoboBrain at Cornell University and Stanford, which involves applying machine learning methods to robots, says BrainOS could appeal to those seeking to commercialize new kinds of robots. “There is a need for an easy-to-use product that combines higher-level skills such as vision and motion planning,” he says.
Brain Corporation is also releasing a version of its software with a chip developed by Qualcomm called bStem (short for “brain stem”). This chip is also designed to mimic the way the brain works, storing and processing data in parallel rather than in series. Such “neuromorphic” chips can be used to run simulated neural networks very efficiently, and Qualcomm is one of several companies hoping to commercialize the technology (see “10 Breakthrough Technologies: Neuromorphic Chips”).
Robot technology is advancing quickly, partly thanks to advances in hardware including computer chips, sensors, and actuators. But software is also driving progress. For example, the open source Robot Operating System makes it easier for engineers to add new capabilities to a robot without having to build basic functionality from scratch.
Advanced machine-learning software such as BrainOS could see other, more advanced capabilities shared between robots, with robots learning from each other’s experiences.