The robotics journal Autonomous Robots has its own blog, which is intended to take the hardcore robot news from the journal and make it a bit more reader friendly. They also link back to the journal articles, should you need a little of that hardcore techy info. Yeah baby. Anyway, looking back through some of their posts, I found these vids of LittleDog exhibiting some new learning behaviors.

While it’s best to develop a robot that’s capable of adapting to new situations, after the robot explores the world for a bit, new situations (in the general sense) stop showing up as often. So ideally, you want a robot that’s able to recognize a situation that it’s already experienced, and apply those past experiences without having to start figuring out what to do from scratch. To this end, Martin Stolle and Christopher Atkeson are developing a system that allows a robot like LittleDog to first recognize features that it’s seen before (like walls or gaps), and then access a library containing a sequence of actions that it has successfully applied to solve similar situations in the past. LittleDog can then alter those actions to apply them to the current situation rather than attempting to compute an entirely new action sequence. It’s faster, more efficient, and I imagine more successful since LittleDog is building on past experiences that worked as opposed to trying something entirely new.

After the jump, another LittleDog video, just because. (Read more…)

Our last LittleDog update was back in September, when we watched the robot walking on posts and doing dramatic death dives. In this latest version (in terms of software, I assume) of LittleDog from MIT, the robot moves noticeably faster and smoother, and seems more reliable overall, especially over rough and unfamiliar terrain.

It’s interesting to see how LittleDog’s ‘brain’ has been improved… The way that the robot chooses foot placement, for example, uses generalized templates of safe foot positions to determine foot placement on unfamiliar rocks, which is a process that we humans can relate to. Your head contains a database of rock templates, built up from your experiences with rocks. If you’re trying to figure out where to safely stand on an unfamiliar rock, you’re probably thinking about what other rocks you’ve stood on before and comparing them to the current rock to decide whether or not you can stand on it. This may not be an entirely conscious process, but humans (and other animals) learn from experience, and use generalized experiences to help decide what to do in specific instances in the present, and the consequences of those instances help to refine the generalized experience and increase our overall robustness. So anyway, this is basically what LittleDog is now doing. Pretty cool.

As far as I know, while all of LittleDog’s behaviors are autonomous, it does depend on an overhead view of its path, provided by an external camera mounted above the course. Finding a way to integrate sensors capable of providing the information that LittleDog needs to decide where to place its feet is probably a project that’s entirely separate from what MIT’s Computational Learning and Motor Control Lab is trying to accomplish, but now that LittleDog has gotten so good at the whole walking over rough terrain thing, making the robot self sufficient seems like it could be an obvious next step, as it were.

[ USC CLMC ] VIA [ Gizmodo ]

Browsing around the MIT website (as I do from time to time so that you don’t have to), I ran across the web page of Katie Byl (that’s “bill”), who is currently working at the Harvard Microrobotics Lab (on stuff like this). Katie used to study legged locomotion at MIT, which involved developing dynamic motions for LittleDog. And sometimes, these motions didn’t quite work out as planned… Also in the following video are clips of LittleDog walking on pegs (“Karate Kid” style), and a slow motion clip of LittleDog bounding up onto terrain:

Incidentally, the reason why LittleDog collapses the way it does is that DARPA gives points for LittleDog’s speed (irrespective of whether it’s upright or not), and LittleDog can increase its speed over a given distance slight by collapsing forward and then stopping the clock. Katie calls this a “dramatic death-dive.” Dramatic indeed.

[ Robot Locomotion @ MIT ]
[ Katie Byl @ MIT ]

Researchers at places like MIT have been using Boston Dynamics‘ LittleDog robot for years now as a testbed to teach legged robots to learn how to traverse variable terrain on their own. This video shows some highlights of a “dynamic double-support gait,” which means (as near as I can tell) that LittleDog is supporting itself, at times, on only two of its four legs. This is a substantially more efficient way of negotiating terrain than we first saw two years ago. LittleDog also demonstrates some markedly biological ways of negotiating obstacles (with the possible exception of the belly flop on the Jersey barrier)… I especially liked how it pranced in place slightly before tackling each stair. All this stuff is obviously a lot of work for a little bot, since poor LittleDog completely collapses at the end of every test.

LittleDog, remember, is teaching itself the most efficient way to negotiate these surfaces. Overhead cameras examine the terrain and plan out LittleDog’s route by computing a ‘cost’ for each step, which takes into account the distance moved towards the goal as well as the potential for a fall. After a lot of trial and error, LittleDog figures out how to best compromise between progress and stability, and the lessons it learns could be propagated up to other, larger quadruped robots.

This video is from Phase 2 of DARPA’s Learning Locomotion program… MIT’s LittleDog team was awarded funding for Phase 3 of this program back in 2008, so we’ll keep you updated.

[ MIT Robot Locomotion Group ]
[ 2006 Abstract ]

By Evan Ackerman

Robots are notoriously bad when it comes to the unpredictable. This is unfortunate, since most of the environments on Earth are more or less unpredictable. Professor Stefan Schaal from USC Viterbi is attempting to build a robot dog that can autonomously navigate over very rough terrain, and has received $1.5 million from DARPA‘s Learning Locomotion program in order to do so. DARPA is interested because a pack of robot dogs would be great at carrying supplies for troops, as long as they’re capable of keeping up on their own. Currently, they have large dogbots in active development that can handle flats and ramps, but nothing more challenging.

Schaal’s LittleDog is designed more as a software and sensor research platform than to be a hardware prototype for DARPA. LittleDog itself is built by Boston Dynamics, and is fairly complicated. Each of LittleDog’s legs is powered by three electric motors, giving them a lot of flexibility. Sensors measure the angles of each joint as well as body orientation and foot/ground contact. The basic strategy for walking over both smooth and rough terrain is “to adjust a smooth walking pattern generator with the selection of every foot placement such that the center of gravity : follows a stable trajectory. To do this, the robot calculates where and how it should proceed, based on the current position, velocity, and acceleration of its legs. If one effort fails, the dog learns from its mistakes and tries another route the next time.”

There are two clips below; the first is an earlier (wired) version of LittleDog dealing with some pretty rocky terrain, and the second clip shows a wireless (and noticeably more evolved) version.

[ Little Dog ] VIA [ USC Viterbi ]