Cheap Four-fingered Robot Hand Edges Closer To Human Dexterity 73
ananyo writes "A robot that can reproduce the dexterity of the human hand remains a dream of the bioengineering profession. One new approach to achieving this goal avoids trying to replicate the intricacy of the bones, joints and ligaments that produce our most basic gestures. A Sandia National Laboratories research team has adopted just such a strategy by designing a modular, plastic proto-hand whose electronics system is largely made from parts found in cell phones. The Sandia Hand can still perform with a high level of finesse for a robot, and is even capable of replacing the batteries in a small flashlight. It is expected to cost about $10,000, a fraction of the $250,000 price tag for a state-of-the-art robot hand today. The Sandia Hand's fingers are modular and affixed to the hand frame via magnets. This gives the researchers the flexibility to design interchangeable appendages tipped with screwdrivers, flashlights, cameras and other tools. The fingers are also designed to detach automatically to avoid damage if the hand hits a wall or other solid object too hard. The researchers say the hand can even be manipulated to retrieve and reattach a fallen finger. The Hand's current incarnation has only four fingers, including the equivalent of an opposable thumb. In the video with the article, the Sandia Hand demonstrates a number of capabilities, including, perhaps most impressively, dropping a AA battery into a flashlight."
Re:Why is this hard? (Score:5, Interesting)
I'm surprised that this and walking are such difficult tasks for robots. I would have thought that reverse engineering the hand would be easy once you've got actuators working. And the human gait has been observed to death and yet we can barely get the robots to walk. It's amazing that these structures we have working examples of cannot be mimiced yet in this day and age. Working consciousness, computer vision, anything that involves some sort of understanding on the part of the machine - I get. But a physical thing like the hand or the human gait? Both seem really well understood.
But I guess they apparently aren't.
Locomotion is reasonably well understood now, but that took a long time. The posting above illustrates one of the major misconceptions. Locomotion is not about gaits. For over a century, starting when Muybridge took the first movie, people did gait studies and obsessed on footfall patterns. That's all wrong.
The first big breakthrough was when Raibert built a self-balancing one-legged hopper. With one leg, there aren't many gait options, and balance dominates the problem. Basic balance on the flat is 1) when in contact with the ground, level the body, and 2) when in the air, position the foot for a landing at the point that will result in zero change in speed. Displace the landing point slightly to accelerate or decelerate.
On the flat, it's all about balance. Once you get off flat surfaces, traction control starts to dominate the problem. I did some work on that. It's like ABS for feet; the robot must keep side forces below the break-loose point.
Once you have basic traction and balance, gait is an emergent behavior. Which foot can most usefully achieve the traction and balance goals? (With more than 4 legs, there are many options.) When something is maneuvering fast or recovering from an upset, there's concept of repetitive "gait". It's more of an asset management problem. Look at some of the Big Dog videos in detail to see this.
Robot manipulation has been underestimated for decades, too. McCarthy once thought it was going to be a summer project to program a robot to assemble a Heathkit TV set kit. Big underestimation. That's still beyond the state of the art. (Stanford actually bought the TV set kit, which was finally assembled by some student and put in the CS department lounge.)
Robots manipulate all sort of useful things in controlled environments, but manipulation in uncontrolled environments is still very poor. Willow Robotics has demonstrated towel and sock folding, which is cutting-edge work. The DARPA ARM program, not so much.
Progress is picking up now that enormous compute resources can be devoted to the problem. It will pick up further when a simulator good enough to debug in is developed. DARPA is funding Willow Garage to upgrade Gazebo to do that. I suspect that the physics engine Gazebo uses is not up to the job, but that can be fixed by applying enough money.
Money is important here. We're now at the point where throwing money at robotics produces real progress. That wasn't true 20 years ago, when NASA blew something like $100 million on the Flight Telerobotic Servicer and got zip.