Created by Boston Dynamics at request from DARPA, shortly after the BigDog pack mule prototype was created. LittleDog is a timid-looking four-legged robot about the size of a Chihuahua.

It is a research robot, designed to be used to find better locomotive learning approaches, which can then be transferred to its larger cousin. A standardized hardware configuration, with a carry handle for convenience, LittleDog was involved in a three stage test by DARPA and a half dozen leading AI labs, in order to probe the fundamental relationships among motor learning, dynamic control, perception of the environment, and rough-terrain locomotion.

In other words, to create robots that teach themselves how to walk, are aware of their environment, and can adapt to it, based on their own past experiences. Because the robotic platform is standardized, and the only changes are in software - the mind of the robot - it is relatively easy to take parts of the work of one lab, and combine it with work from another, then expect it to work as a new robot locomotion mind.

The robot has three motored joints on each leg, and its movements are controlled precisely by an on-board computer. An internal gyroscope lets the robot sense its orientation, while an external motion-capture system monitors the precise position of each limb and joint as it moves - proprioception for the robot.

The Phases of the Learning Locomotion Project

Phase I

Phase I lasted 15 months, including time for development of learning software, which interfaced with the hardware via a standard API, in ROM, that Boston Dynamics provided. Each lab had to create software capable of learning how to walk, then learning how to walk on uneven, unpredictable terrain.

The goal speed was set at 0.6 in/sec and scale obstacles up to 2.5 in tall - half the leg height of the robot.

This phase completed successfully, in 2006, with all six labs managing to achieve the target.

Phase II

Phase II was slightly longer, an 18-month effort that concluded in early 2008, in which the robots were taught to move much faster - 3.8 in/sec, with the max obstacle height being 5.7 in, or, more than the height of the legs of the robot.

The speed might still seem slow, but for a robot only 11 inches long and with limited power, it is still quite a pace. Again, six labs completed the challenge, with a variety of different approaches. Some of the systems, such as the Stanford entry in the video below, had the LittleDog slowly teaching itself how to walk, and crawling like a newborn, unsteady on its feet and falling over frequently. This newborn learnt quickly, and as you can see, by the end of the video some months later, was fully confident on any terrain.

Other approaches put balance above locomotion - the robot had to remain perfectly balanced for its load, before it even thought about moving.

Phase III

Phase III, currently under way, is a series of tests, that started summer 2008, designed to see how the robot did against obstacles it had never been programmed for. In the video below, an MIT LittleDog encounters stairs for the first time, and has an unauthodox method of tackling them.

Phase III is technically complete at time of writing, but the results have yet to be published by DARPA, or by any of the six labs (excluding the rare, occasional video).Thus, very little information is available at this time.

Universities Involved:

Boston Dynamics (Carnegie Mellon University)

Institute for Human and Machine Cognition (IHMC)

MIT

Stanford

University of Manchester

University of Zurich

References

Local

Robot dogs race to be soldier's best friend

BigDog

External

Boston Dynamics: LittleDog - The Legged Locomotion Learning Robot

Metastable Legged-Robot Locomotion (PDF)

IPTO :: Programs :: Learning Locomotion (LL)