Undergraduate Research Opportunities in the Atkeson Lab

I am looking for undergrads to develop controllers for humanoid robots, especially for walking. Garth Zeglin has built a series of smaller bipeds for this research, and we are getting a bigger biped soon from Sarcos. We will explore several different approaches (see below). These approaches look like optimal control approaches, but become learning approaches when a model must be learned as the robot attempts the task. Learning also plays a role in speeding up this type of planning, and leading to better plans.

Plan A: Policy Search

Policy search is currently one of the "hot" approaches to robot learning. A controller is manually defined, and then automatic algorithms are used to find good parameters for the controller. Some papers you can read about this:

Inverted autonomous helicopter flight via reinforcement learning, Andrew Y. Ng, Adam Coates, Mark Diel, Varun Ganapathi, Jamie Schulte, Ben Tse, Eric Berger and Eric Liang. In International Symposium on Experimental Robotics, 2004.
Rhex. Look in IEEEXplore for Weingarten papers. There should be an ICRA03 paper which is better than the ICRA04 paper listed here.
German Aibo Robo Soccer
CMU Aibo Robo Soccer
Texas Aibo Robo Soccer
English Aibo Robo Soccer
Automatic gait Optimisation for Quadruped Robots M Kim, W Uther Proceedings of 2003 Australasian conference on robotics and automation
Hornby's stuff

Numerical Recipes is online (free) and has a nice explanation of function optimization techniques. Matlab has an optimization toolbox that is useful as well. The GNU Scientific Library has optimization code. Genetic algorithms code is widely available on the Web.

Here is a biped simulator written in C, with graphics for the X11 windows system (so it is easy to run it in Unix/Linux, but needs some conversion to run it in Windows): Zip format
Look in biped/biped/notes to get some instructions. It uses the Numerical Recipes in C library, which you must provide.

Plan B: Full Dimensional Dynamic Programming

Remi Munos has developed some nice algorithms for large scale dynamic programming. I would like someone to implement his algorithms, first on his examples and then on the biped.

Plan C: Reduced Dimensional Dynamic Programming

Mike Stilman developed a planning approach that used simplified models. This approach needs to be tested more thoroughly, and then applied to actual bipeds.

Plan D: Trajectory Optimization

We will try applying several trajectory approaches. DIRCOLhas been used to optimize biped walking.

Plan E: Reduced Dimensional Trajectory Optimization

An interesting example of reduced dimensional trajectory optimization.