Global sales of robots have doubled in the last five years, but the robots deployed in factories and warehouses today are practically the same as we had decades ago. They are powerful and precise, but costly to buy and dangerous to the humans next door.
Blue, a new UC Berkeley robot, aims to break that mold with the help of AI.
The blue looks a bit like the drawing of a child's robot: it is made of bulky pieces, printed in 3D, and has a pair of humanoid robot arms with tweezers for the hands. It can be controlled using VR phones, which allow operators to move their arms and then Blue will shake his altogether. You can also train to manipulate objects using artificial intelligence, a control method that is still surprisingly rare in robots.
Pieter Abbeel, the robotist who leads the project, wants to change this, and says that Blue has been built from scratch to take advantage of the recent improvements in AI. "The fact that artificial intelligence is increasingly capable gave us the opportunity to rethink how to design a robot," says Abbeel The Verge .
Abbeel explains that most of the robots in use today are designed to be strong and precise. Their movements are predefined and they simply repeat the same action over and over again, either lifting pallets of goods, welding cars or fixing screws on a smartphone.
The robots of the future, in comparison, will be reactive and dynamic. They will be able to work safely with humans without crushing them, and instead of planning their actions in advance, they will navigate the world in real time using cameras and sensors.
"If you look at traditional robots, they are designed around the principle of very high precision and repeated movements." Abbeel says. "But it does not necessarily need a submillimeter repeatability." (That is the ability to perform the same task over and over again with differences in movement of less than one millimeter). "Human beings do not have submillimeter repeatability, instead we use our eyes and our sense of touch to do things through feedback."
Abbeel and his team, researcher Stephen McKinley and graduate student David Gealy, expect Blue to work the same way. It has a central vision module with a depth detection camera, and its arms are controlled by motors with rubber bands that give it flexibility. If you press against an industrial robot arm, it's like pushing against a brick wall. But Blue is more like a human in a subway car full of people: push it and it will move to the side.
This makes Blue safer to work, but it is also suitable for research that uses reinforcement learning, a type of AI training method that is becoming popular in robotics. Reinforcement learning works by asking an agent to complete a task and rewarding it when it does. Basically it is trial and error, since the agent starts without knowing how to complete his goal and then slowly teaches over time.
Using traditional robots with reinforcement learning can be expensive. Their lack of flexibility makes them brittle and prone to damage. In addition, reinforcement learning takes time to produce results, and since robots are expensive, costs accumulate quickly.
This is another area where Blue can make a difference. PR2, a popular research robot built by Willow Garage that also has a pair of arms and clamps, makes researchers earn around $ 400,000. The list of materials for Blue, in comparison, is only $ 3,000. Abbeel says the team has not decided on a final price, but they hope to reach the $ 5,000 range.
"That's possible when you're willing to give up submillimeter accuracy because you realize you do not need it with AI-based control," says Abbeel.
Many other research labs and startups are also addressing this new paradigm, hoping to teach robots to work using artificial intelligence. Abbeel is the president of one of them, a startup called Embodied Intelligence. Kindred AI, a company that builds robots that collect items in warehouses, is another. OpenAI, the research laboratory founded by Elon Musk, has done similar work with robot hands, and Google is also exploring AI training for robots.
Still, some experts are skeptical about Blue's appeal. They point out that it is not so different from Baxter, another robot with arms and tweezers that was destined to work alongside humans. The company that Baxter did, Rethink Robotics, closed last year.
Ankur Handa, a robotics researcher at Nvidia, said Blue's clamps limit the range of tasks he can perform, and his lack of precision would be a problem, even with AI controls. "In general, I do not think they offer anything new," says Handa The Verge .
But Abbeel is optimistic about Blue's future. The robot is being built in small lots at this time, but Abbeel hopes to scale, eventually moving to outsourced manufacturing to produce larger numbers. The first target customers will be research laboratories and universities where robots are currently shared between teams, like computers in the 1960s. Offering a cheaper robot will make them more accessible, which will increase the production of research of robots.
More importantly, Abbeel expects Blue to provide a blueprint for what could be the domestic robot of the future: something that is inexpensive, flexible, and plays well with humans. "Home is absolutely what we have in mind with this type of design," he says. "There are still many challenges ahead, and this is not how we think this specific robot is going to a home. [But] this is a design paradigm that takes us in a new direction."