Creating robots that think for themselves

From Earth, you want to pick something up from the surface of the moon. Using a joystick, you control a grabber — but you keep missing. It's the delay: It takes light nearly three seconds to get to the Moon and back, and the control signal even longer.

The big picture: This is a fundamental, insurmountable hurdle to moving stuff around in space if you're not there. The farther you are from Earth, the longer the communications delay — nothing, after all, moves faster than light.

• One workaround is to just move really slowly. Fine if you're picking up a rock; not that great if you're trying to build a structure or repair something from afar.
• A more powerful approach is to build as much autonomy as possible into the robot doing the job. That means that instead of telling it to move the grabber forward two inches, down four inches, and grab, you can just tell it to pick up the rock.

What's happening: Several startups, as well as NASA, are trying to find the right combination of human control and autonomy that would allow robots to do complex tasks in space. Full autonomy isn't an option for now — today's artificial intelligence isn't up to the job yet.

• NASA wants to refuel or fix up orbiting satellites that weren't designed to be serviced. So it's developing robots that autonomously find and clamp onto satellites, at which point technicians on the ground use joysticks to carefully remove a covering and a valve, and start pumping gas.
• SE4, a startup in Tokyo, has a system that blends human control and autonomy. An operator wears a VR headset and manipulates a digital scene from a robot's-eye-view. Then, the robot does the same thing — but not too precisely.

How it works: The SE4 system, which is being demonstrated today at the SIGGRAPH conference in Los Angeles, abstracts away some of the small details of the operator's actions, focusing instead largely on the intention.

• Imagine someone demonstrating how to build an IKEA dresser. They won't try to tell you how many degrees to move your elbow before tightening your fingers.
• Instead, they'll call out bigger-picture actions, like "align these two parts" and "now screw this in." That's what the robot does, too, having pre-learned basic actions like aligning and screwing in.

"We want human minds to ride these robots into remote locations and learn things," says Lochlainn Wilson, CEO of SE4. "We could have them going outside and doing operations without endangering human lives, yet humans will be gaining knowledge and experience."

What's next: By 2022, SE4 wants to remotely build a lunar base — somewhere on Earth. It'll be a proof of concept for its eventual aspirations to do construction on the actual Moon.