Learning Locomotion: Extreme Learning For Extreme Terrain

Chris Atkeson, Drew Bagnell, James Kuffner, Matt Zucker, Nathan Ratliff

This page describes CMU's participation in the DARPA Learning Locomotion Program. This program involves using a quadruped robot constructed by BDI, Little Dog, to explore learning to be more agile and robust. There are a series of tests our robot must perform. Some videos and information on these tests as they occur are given below.

PI meeting slides, 8/15/06

PI meeting slides

Test 0: Walking on Flat Ground, 4/18/06

Start in a random orientation, locomote to a goal position, stop. Here is a video of a practice run at CMU of our trot: MPG format or AVI format

Test 1: Crossing the Mesa, 5/2/06

Cross over a step obstacle (step up followed by step down). Here is a video of a practice run at CMU: MPG format or AVI format
Here we walk half on the sideways version of the obstacle. MPG format or AVI format

Test 2: Rocks and Mesa, 6/6/06

Cross over a pile of flat rocks and a step obstacle. Here is a video of a practice run at CMU: MPG format or AVI format

Test 3: Tilted Terrain, 7/6/06

Cross over a tilted pile of flat rocks and a step obstacle. Here are videos of practice runs at CMU:
run 1 MPG format or AVI format
run 2 MPG format or AVI format
run 3 MPG format or AVI format
Here are more videos of practice runs at CMU where we tried other terrain board tilts: expected layout, last board tilted up and to the side, add spacer, uphill-downhill, another uphill-downhill with step obstacle, uphill-downhill with rocks at end, and oops.

Test 4: Tilted Terrain, 8/15/06

Same terrain as Test 3.

Test 4 practice data

For all of us to learn to handle variations among dogs better it would be useful to have data from all dogs. My suggestion is to make data from recent practice trials (say Test 4 practices, so the terrain is similar) available to everyone. I put 25 data files on Test 4 terrain in a zip file available from www.cs.cmu.edu/~cga/leg-learn/ It would be ideal if all the data was from batteries, so the characteristics of the power supply are more similar, but runs using an external power supply are also useful. Just let us know.

Under battery power: t4practice.zip

Test 4 actual data

We should be able to tell the difference between the first 3 CMU runs in Test 4 (LLGT Dog 1, did badly) and the last 2 (LLGT Dog 2, did better). [CMU Data] We should be able to tell the difference between the last IHMC run in Test 4 (LLGT Dog A, did well) and all previous IHMC runs (LLGT Dog B, did badly). [IHMC data] Note that we didn't keep track of whether LLGT Dog A = LLGT Dog 1 or LLGT Dog A = LLGT Dog 2.

Test 5: Terrain O' Death, 9/7/06

All we had to work with are these photos, so no practice runs.

Test 6: Return of Terrain O' Death, 10/5/06

Same terrain as Test 5. Here are some videos of practice runs at CMU: travelling east and west. Human drivers have found two routes travelling north, but we have not yet got the dog to do either of them autonomously. Here is the dog trying one [video]. We have not yet found a route travelling south [video]..

Here are some papers relevant to our work for Test 6: Learning From Demonstration (IROS06) and Trajectory Library (ICRA06).
Martin Stolle's thesis proposal.

Test 7: Son of Terrain O' Death, 11/1/06

Same terrain as Test 5, except now we go East. The big issue this month is ROBUSTNESS. How do we get good performance at CMU to transfer to the test site? One part of the answer is testing with simulated bad conditions.

Here are some videos of practice runs at CMU:
1) Starting on the rocks, wearing slippery booties, and having an off center load strapped to the back of the dog: [MP4 video]. [MP2 video].
2) Varying start board height: test 6 height, test 6 height + 0.5 board width, test 6 height + 1 board width. [video].
3) Our fastest run so far, almost 2cm/sec across the terrain board. [video].
Here are the planned footstep locations for the test (used in the fastest run video above). Movement goes from left to right (towards East), red is the front right foot, magenta is the front left foot, black is the rear left foot, and blue is the rear right foot.

Here are the footstep and body locations at touchdown for the plan, and here is a program to parse that file.

Test 8: Blind Test, 11/1/06

We didn't know in advance what this test would be, so no practice videos.

Test 9: Christmas Test, 12/20/06

The new terrain turned out to be similar to Terrain D, so we practiced using Terrain D (aka Terrain 'O Death).

Test 10: New Year's Test, 1/17/07

For some reason we did not videotape our practice runs on Terrain G.

Test 11: Blind Test, 2/8/07

Multiple terrain boards with step, using planning: MP4 or MP2

Handling no-go gaps between boards, using planning: MP4 or MP2

A nice video of learning from demonstration running on Terrain D. MP4 or MP2
Short version of above: MP4 or MP2

Test 12: Blind Test, 2/22/07

Before and After Learning Video: Comparison of planning performance with human generated terrain cost function and learned cost function on Test 11A terrain: more compressed version or less compressed version

LFD Generalization Video: Comparison of learning with demonstration using terrain coordinates with learning from demonstration using local terrain features and planning to choose which learned steps to use and to fill in gaps. The dog's goal is to get totally on to the 2nd rock board. The dog has not previously experienced the transition between boards. MP4 or MP2

How fast can the dog go?: A speed test we did a while ago under external power (20cm/sec). We will retest using battery power, but our recollection is something like 14cm/sec under battery power. We don't think the dog can go anywhere near this speed over rough terrain, due to joint speed limitations. video

Phase II goal

Our Phase II goal. (We don't know who made this video, so we can't give proper attribution or credit. It is very impressive).