Feeds:
Posts
Comments

Archive for the ‘new-stuff’ Category

BigDog

Of course, if we talk about LittleDog, we should look at the BigDog video.  All automated quadruped.  Some look at it as scary, others might think this is another drain of the world’s energy, … At the same time, what marvelous things man creates.  BigDog:

Copyright (c) 2008 by Waiming Mok

Advertisements

Read Full Post »

LittleDog Learning to Walk

At the RAS meeting, Zico Kolter presented his work on learning algorithms for LittleDog (a quadruped) for path planning and control, as well as some initial work on perception.   In path planning, the problem is framed as solving a convex optimization problem on footsteps and foot placement.  The human provides advice on where Littledog should step when it makes a mistake.  Littledog translates the advice to adjusting parameters in the optimization problem.  With 10 minutes of training, Littledog can plan and walk across new terrains the it has not seen before.

On control, the problem is framed as making LittleDog learn to walk fast on a specific complicated path with specific orientations.  Walking fast involves having 2 diagonally opposed legs up in the air at each step.  After 100 simulations by varying some parameters, the learning algorithm based on SVM reduces the problem from 98-dimension to 4-dimensions.  This takes 2 minutes.  LittleDog is then able to manueuver thru the path.

These algorithms helped Stanford win the DARPA challenge.

Copyright (c) 2008 by Waiming Mok

Read Full Post »

« Newer Posts