June 20, 2019
A new chip lets robots “imagine” their actions before they make a move
Robots that can rapidly plan out their movements could accelerate factory automation—and help keep fragile humans safe.
“Even if you’re not worried about having humans next to the robot, you might want to modify your cell without incurring the cost of bringing in a technician,” says Sean Murray, a robotics engineer and cofounder at Realtime Robotics who showed me around.
The movement problem
A number of companies are trying to find ways around this problem. Some are testing sensors that will stop a powerful robots in its tracks if it spots an obstacle. Realtime Robotics is trying to go further, by giving robots the kind of low-level intelligence needed to move through the real world. This is the physical awareness that humans and animals take for granted whenever they move an arm or a leg.
In several different rooms at Realtime, industrial robot arms are testing the capabilities of a new chip that the company has developed to make this possible. When hooked up to 3D sensors, this chip lets the machines rapidly consider a range of different actions, effectively “imagining” the outcome, before choosing the one best suited to the task at hand. In one room, I watched as two robots performed balletic feats of teamwork, gliding around one another and occasionally handing over items.
“The fundamental challenge is that robots are so stupid,” says George Konidaris, founder and chief roboticist at Realtime as well as an assistant professor at Brown University in Providence, Rhode Island. “We have this basic motor competence and robots don’t.”
Motion planning is deceptively difficult for a robot, partly because each joint adds an extra dimension to the calculations that must be performed.
Make your move
The company’s chip supercharges the mathematical computations behind a relatively simple motion-planning algorithm developed by Konidaris and others while he was at Duke University. By running the computations in parallel, the dedicated chip can perform them more than 10,000 times more quickly than a regular computer chip, while also using less power.
“The approach is very clever,” says Tomás Lozano-Pérez, a professor at MIT who advised Konidaris when he was a graduate student.
[Originally posted by MIT Technology Review — June 17, 2019]