Correcting robot mistakes using mind control without thinking


When it comes to controlling robots, it isn’t just a matter of finding ways to give them commands, but of making sure they’re carrying out those commands properly. To help with this, a team from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and Boston University has come up with a system that turns the operator into a human automatic fault detector to alert the robot when it makes a mistake.

Mind-controlled robots aren’t new, but they’re also still very much in the developmental phase. Direct control from brain to machine may do away with the problems of a mechanical interface or teaching a robot to respond to voice commands, but by itself it isn’t enough. It’s one thing to order a robot to do this or that, but it’s another thing entirely to get to do it right.

According to MIT, past work on controlling robots has involved hooking up the operator to an electroencephalography (EEG) monitor, then teaching them to give the robot orders by thinking in certain, carefully prescribed ways, like looking at one of two bright light displays to tell the robot which task to perform. The problem is that it was very much a one-way process and exhausting because it required constant attention.The CSAIL approach was to make the command path into more like a feedback loop by monitoring the operator’s brain in such a way that they tell the robot in real time when it’s making a mistake without the operator doing anything.

“Imagine being able to instantaneously tell a robot to do a certain action, without needing to type a command, push a button or even say a word,” says CSAIL Director Daniela Rus. “A streamlined approach like that would improve our abilities to supervise factory robots, driverless cars, and other technologies we haven’t even invented yet.”

The idea is to look not at the brain’s conscious commands, but at what the team calls “error-related potentials” (ErrPs). These are signal patterns that the brain puts out when a person notices a mistake. In other words, the operator can warn the robot when it’s doing something wrong without consciously thinking about it. The signal is sent automatically to the robot and it takes on the burden of learning instead of the human.To develop this system, the team used a Baxter robot from Rethink Robotics to carry out simple sorting tasks under the instructions of a human operator wearing an EEG cap. Currently the system is limited to dealing with binary-choice activities, but running the signals through an algorithm saw the received brainwaves processed in about 10 to 30 milliseconds.

The CSAIL team says that the ErrP signals are extremely faint, so the feedback loop needed some tweaking to get the right results. This involved making sure the signals were properly classified and by monitoring what are called “secondary errors,” where the robot fails to respond to the first error signal. When this happens, the brain provides additional reinforcement that improves accuracy, which the team hopes will reach as high as 95 percent once the system can recognize secondary errors in real time. The team also says that ErrP signals increase proportionally to the size of the robot’s mistake, which could lead to future systems able to deal with more complex multiple-choice tasks.

In addition to controlling robots, the team hopes that the technology will also be of benefit to people who lack the ability to communicate verbally.

The study results are available here (PDF).

The video below gives an overview of the brain-controlled robot technology