Geo-II was my second attempt at building a flexible-spined walking machine. GEO-II is much lighter that GEO-I. In addition, we have force sensitive feet. GEO-II is currently being used in conjunction with an evolving neural controller (see the lamprey).






This is GEO-I. It was built in the summer of 1994 at USC. The key innovations of GEO-I are A) a flexible spine and B) reconfigurable legs. The reconfigurable legs and flexible spine allow the robot to walk with outspread legs in a salamander-like gait. When the legs are tucked under the body, the robot can walk with a cat-like or mammalian gait. This was designed and built under contract to JPL. GEO-I sports a TI TMS320C30 DSP controller. Currently, GEO-I resides at SRI International

SRI International home page















DOWNLOAD QTmovie Rodney walking 2.2 M

DOWNLOAD QTmovie Rodney learning to walk 2.5 M





Rodney was built in 1992 at USC. Three of us participated in its construction: Andrew H. Fagg and Arvin Agah and myself. Andy and I wrote a couple of papers [1,2] describing how we evolved a neural network, using GAs, to control the walking of this machine.

All evaluations were done on the real machine (not in simulation). The key idea was to decompose the "learning to walk" into two functional stages. I had this idea after I read Coghill's wonderful book: "Anatomy and the Problem of Behavior" at Michael Arbib's suggestion. In this book Coghill describes the development of a salamander embryo. What we see during development is that the animal develops certain gross behavior which is refined in several clearly definable stages. We used this inspiration for the idea of "staged evolution" of Rodney's behavior.

In the video, you will see the Rodney in several "snap shots" during its last stage of "evolution." Here Rodney struggles forward, then finds a swimming gait, and finally settles on a tripod gait. During the tripod gait we see Rodney being "teleoperated" by a remote operator. The remote operator activates Rodney's neural network and causes it to move forward, backward and to turn. To my knowledge this was the first time that GA's evolved walking in a real robot.

[1] M. Anthony Lewis and Lucia S. Simo, "Neurocore: Evolution, Development, and Robotics" Sumitted to IROS 96, Osaka

[2] M. Anthony Lewis, Andrew H. Fagg, and George Bekey, Genetic Algorithms for Gait Synthesis in a Hexapod Robot, in Zheng, ed. Recent Trends in Mobile Robots, pp 317-331, World Scientific, New Jersey, 1994.

[3] M. Anthony Lewis, Andrew H. Fagg, and Alan Solidum, Genetic Programming Approach to the Construction of a Neural Network fo Control of a Walking Robot. in Proceedings of the 1992 IEEE International Conference on Robotics and Automation, Nice France, pp 2618-23, 1992.