Last week at the Hot Chips 23 conference held at Stanford, Willow Garage President and CEO Steve Cousins gave attendees a 1-hour taste for the difficulties associated with designing autonomous robots that can freely interact with people without hurting them. Despite the amazing things a Willow Garage PR2 robot can already do—like play pool and make pancakes—there are still many, many things the robot cannot do well. For example, getting the robot to roll out on stage during the keynote proved far more difficult than you might expect. Watch:
As Cousins explained, robotics is not yet a large enough market to support the development of robot-specific sensors so Willow Garage and other robotics pioneers make do with sensors developed for other markets. For example, the “beak” that you see moving up and down where the PR2 robot’s “mouth” might be expected is a low-cost laser scanner developed for the garage-door industry. Lately, PR2 robots have been wearing Microsoft Kinect 3D spatial sensors as “hats” because there’s an incredible amount of technology packed into this device that sells for less than $150 in unit quantities. The Microsoft Kinect, said Cousins, is an excellent 3D spatial mapper at “couch” distances but anything farther away than about six to eight feet merely registers as “far away.” Here’s a photo of the PR2 at the Hot Chips 23 conference showing these sensors.
Robots also need a lot of distributed processing as it turns out. Local sensor processors for touch and local actuator processors for motors would be extremely helpful said Cousins. He also needs a robust power/signaling standard that will work through slip rings so that the signals can make it all the way to a gripper, for example. FireWire has no chance, he said. USB is OK and Ethernet works great through the intermittent connection of a slip ring.
Here is Cousins wish list:
- Cool chips. Robotic designers need infinite computing capacity for zero power because batteries never last long enough. Any battery power used for computing is not available for motors.
- Shared-memory computing. Perception requires a lot of memory and Cousins says that robotic designers want to store sensor data once, because it consumes a lot of room, and they need to perform a lot of perceptual processing in real time so that the robot does what it’s supposed to do without crashing into walls, furniture, or people.
- Fast communications. The more computation the robotics designer can push into the cloud, says Cousins, the better. However, sensors produce a fire hose of data that must either be processed locally or quickly shipped to the cloud for processing.
- GPUs (graphics processing units) with Ubuntu support. Apparently the Ubuntu variant of Linux is the OS of choice for today’s robotics designers.
- Specialized processors. Sensors require special computational abilities. You can emulates these abilities with a general-purpose processor, but you’ll need more cycles and burn more energy doing so.