Robots may have a knack for superhuman strength and precision, but they still struggle with some basic human tasks such as folding laundry and making a cup of coffee.
Enter Blue, a new low-cost, human-friendly robot conceived and built by a team of researchers at the University of California, Berkeley. It was designed to take advantage of recent advances in artificial intelligence (AI) and deep-reinforcement learning to master intricate human tasks. It is also designed to be affordable and so safe that every artificial intelligence researcher—and eventually every home—could have one.
Blue is the brainchild of Pieter Abbeel, professor of electrical engineering and computer sciences at UC Berkeley, and his two-person team. They hope Blue will accelerate the development of robotics for the home.
“AI has done a lot for existing robots, but we wanted to design a robot that is right for AI,” Abbeel says. “Existing robots are too expensive, not safe around humans, and similarly not safe around themselves. If they learn through trial-and-error, they will easily break themselves. We wanted to create a new robot that is right for the AI age rather than for the high-precision, sub-millimeter, factory-automation age.”
Over the past 10 years, Abbeel has pioneered deep reinforcement-learning algorithms that help robots learn by trial-and-error or by being guided by a human like a puppet. He developed these algorithms using robots built by outside companies, which market them for tens of thousands of dollars.
Blue’s durable plastic parts and high-performance motors cost less than $5,000 to manufacture and assemble. Its arms, each about the size of the average bodybuilder’s, are sensitive to outside forces such as a hand pushing it away. The arms also have rounded edges and minimal pinch points to avoid catching stray fingers. Blue’s arms can be stiff (like a human flexing) or very flexible (like a human relaxing), or anything in between.
Currently, the team is building 10 arms in-house to distribute to select early adopters. They are continuing to investigate Blue’s durability and tackle the challenge of manufacturing the robot on a larger scale, which will happen through the UC Berkeley spinoff Berkeley Open Arms. Sign-up for people interested in getting access have already started today on that site.
“With a less-expensive-cost robot, every researcher can have their own, and that vision is one of the main driving forces behind this project: getting more research done by having more robots in the world,” says Abbeel.
Robotics traditionally focused on industrial applications, where robots need strength and precision to carry out repetitive tasks perfectly every time. These robots flourish in highly structured, predictable environments—a far cry from the traditional American home, where you might find children, pets, and dirty laundry on the floor.
“We’ve often described these industrial robots as moving statues,” Abbeel says. “They are very rigid, meant to go from point A to point B and back to point A perfectly. But if you command them to go a centimeter past a table or a wall, they are going to smash into the wall and lock up, break themselves, or break the wall.”
If a robot is going to make mistakes and learn by doing in unstructured environments, rigid robots just won’t work. To make experimentation safer, Blue was designed to be force-controlled, which means it is highly sensitive to outside forces and always modulating the amount of force it exerts at any given time.
“With this robot we can make it force-sensitive, nice, and reactive, or we can choose to have it be strong and rigid,” Abbeel says. “Researchers can adjust how stiff the robot is, and what kind of stiffness; do you want it to feel like molasses or like a spring? Or maybe a combination of those? If we want robots to move into homes and perform in unstructured environments, they are going to need that capability.”
To get these capabilities in Blue at low cost, the team considered what features it needed to complete human-centered tasks, and what it could do without. For example, the researchers gave Blue a wide range of motion. It has joints that move in the same directions as a human shoulder, elbow, and wrist so humans can more easily teach it how to complete tricky maneuvers using virtual reality. But the agile robot arms lack some of the strength and precision of typical robots.
“What we realized you don’t need a robot that exerts a specific force all the time, or a specific accuracy for all time, “says Abbeel. “With a little intelligence, you can relax those requirements and let the robot behave more like a human being to accomplish those tasks.”
Blue can continually hold up two pounds of weight with arms fully extended. But unlike traditional robots characterized by one consistent “force/current limit,” Blue is designed to be “thermally-limited,” Abbeel explains. That means that, similar to a human being, it can exert a force well beyond two pounds in a quick burst until its thermal limits are reached, then it needs time to rest or cool down. This is similar to how a human can pick up and carry a laundry basket across a room but might not be able to do so over a mile up a steep hill without frequent breaks.
“Essentially, we can get more out of a weaker robot,” Abbeel says. “And a weaker robot is just safer. The strongest robot is most dangerous. We wanted to design the weakest robot that could still do useful stuff.
“Researchers had been developing AI for existing hardware,” he continues, “and, about three years ago, we began thinking, ‘Maybe we could do it the other way around. Maybe we could think about what hardware we could build to augment AI and work on those two paths together, at the same time.’ And this is a dramatic shift from the way a lot of research has taken place.”