Tuesday Voice: The robot’s challenge

A walk in the woods

By Chuck Smythe

Simply fascinating, the August 18 NY Times science article, “Atlas, a Humanoid Robot, Takes a Walk in the Woods,” by Katie Rogers. The best part of it is the set of videos near the end of the “DARPA Robotics Challenge 2015.” It’s about nine hours worth, much of it as lively as watching paint dry – but nevertheless extremely interesting as engineering, as drama, and as pop philosophy. I encourage readers to skip around and sample a few things to get the flavor.

    The setup: DARPA stands for the Defense Advanced Research Projects Agency, a military operation dedicated to encouraging blue-sky, high-risk research. Their most famous effort, as far as I know (60’s – 70’s), was to fund research to get computers to talk to each other over a phone line. (This has been a little bit important since.) The Robotics Challenge was a competition to build robots capable of operating in field conditions (imagine entering a disabled nuclear reactor to throw some critical switches, or searching a partly collapsed building – but you can let your imagination run wild.)
    The competition required each robot to

1. Drive an ATV through an obstacle course and park in front of a building.
2. Get out of the car without falling down. (One of the most frequently flunked tests!)
3. Walk to the building and find the door.
4. Open the latch, push open the door, and enter the building.
5. Locate a big hydraulic valve and close it one full turn.
6. Locate a board with two big power plugs. Unplug a cable from one outlet and plug it into the other.
7. Select a portable saber saw from a shelf. Use the saw to enlarge a hole in a piece of gypsum board.
8. Head for the exit and find the way blocked by your choice of a) a jumble of concrete blocks, or b) a small pile of construction trash. Surmount the obstacle any way you see fit and leave the building.
9. Climb a four-foot set of stairs.

    The teams were international; contestants were from Korea, Japan, Hong Kong, Germany, MIT, U. of Oregon, Virginia Tech, and several other places. Each robot was closely monitored by a team of eager young men (few women!) with laptops. I don’t think they were allowed to actively control it, though – autonomy was the idea. They were given an hour or so to complete a run of the course. You, naturally, could do the whole thing in ten minutes. The robots, though, had to stop every few seconds to look around and think things over for a minute. This is why much of the footage on the videos is so dull. I wish YouTube provided fast forward.

Some observations:
    Just standing and walking around was a huge problem for most of them. They frequently just fell down and (with one exception) couldn’t get up. Their walk was a timid shuffle that made you want to offer the poor things a walker. Most of them had immense, heavy looking torsos that must have been terribly top-heavy. Driving, on the other hand, was relatively easy.
    Climbing out of the car was so hard that many teams didn’t attempt it, but took a time penalty to hoist their champion out with a crane. Crossing the obstacle was also a big fail for most.

    The saber saw and cable-plugging were also hard, both from a lack of fine motor dexterity and from poor communications with their computers when inside the building.

    Almost all the competitors used vaguely humanoid robots. You know – two legs, two arms, and a head. They then struggled mightily to make their contraptions do all the things we do. You know – walk around, find door handles, and use tools. The champion, though, was a Korean creation that instead asked, “How could we redesign the human body to make all this stuff easier?”

The Korean robot with the saw, and scooting on its knees

Their robots hoisted themselves out of the ATVs by grabbing the roll cage with their hands. Then they fell to their knees, which proved to have wheels that enabled them to confidently scoot around without fear of falling. They stood up when the situation required it, and could walk like all their fellow robots. Faced with the obstacles, they just fell to their knees and bulldozed the trash out of the way. They rolled up to the stairs backwards, rose from their knees by arching far back and giving a wiggle like a Yogi, turned their torsos around 180° so that their eyes were forward and their knees backward, and marched upstairs that way. These are creative people!

    One of the more successful humanoid contestants raised its arms at the top of the stairs and did a victory dance. Then fell down.
    One of the side shows was a mechanical dog. It wasn’t autonomous, but driven by a game controller. There were actually a few of these, all built by MIT, I believe. The most awesome, Cheetah, could run a hurdles course at 30 mph. Others could walk, trot, skip, jump, pirouette in place, sit, and stay. Question: if walking is so easy for these gadgets, why is it so marginal for the general-purpose robots? (It isn’t because the mechanical animals had four legs; there were bipedal walking drones nearly as agile.)

The biggest take-home message, I think, is the demonstration of just how good we meat critters (especially bipeds!) are at the whole balance / agility / proprioception / manual dexterity thing. Decades of engineering have produced only painfully clumsy imitations. Another is that in moving about, we automatically do calculations and make plans at a rate hundreds of times faster than these gadgets with all their banks of computers. How? This is just the latest in a long list of surprises we’ve had about which of our abilities are easy and which are hard. A computer kicked my butt at chess forty years ago, yet going to the fridge for a beer is a desperate challenge even for the latest and greatest. And last I heard, computer language translation is still basically an unsolved problem. We humans (or most any vertebrate) are astonishing beings, and it’ll be a long time before we see The Terminator.
    Last thought: a bunch of nerds sure had themselves a great time.

Copyright © 2015 by Chuck Smythe