 Aim and Introduction From a control theory point of view, anthropomorphic bipedal walking robots might be more difficult to deal with than wheeled robots. The most fundamental aspects for bipedal robots concern the motion behaviors, which allow the robot walk, run, turn, climb stairs etc. The dominating paradigm addressing these issues up to now is based on classical control theory, in which a reference trajectory is specified for the robot to follow. One of the major drawbacks using such methods is that reference trajectories rarely can be specified in a realistic, dynamically changing environment. A bipedal robot will encounter unexpected situations in the real world, which cannot all be accounted for on before hand. The aim of this project is to investigate an alternative approach to address these problems. The methodology is mainly based on evolutionary algorithms (EAs), or more precisely Linear Genetic Programming (LGP). EAs all draw upon the ideas from Darwinian theory on natural selection and survival of the fittest, i.e. they mimic natural evolution in some ways. The long-term goal of this project
is to develop methods for the creation of robust and sustainable
biped locomotion controllers, capable of handling the rapidly changing
dynamic properties of a realistic environment. Simulation Experiments Below are some "sample" animations of successful individuals that evolved. First are individuals evolved with a "minimalistic" LGP. That is, the only feedback that the individuals get when controlling the silmulated biped (apart from fitness of course) are the current joint angles. In each time step of the simulation, those values are fed into the registers of the virtual register machine. The actual individual then use the registers to compute the motor commands for the biped, for the next time step.
Below, the resulting individuals are evolved with registers holding three constant values, which could be used by the individuals to synthesize arbitrary constants. Also, some additional feedback is used to compute the motor commands. The feedback is provided by several virtual Inertial Measurement Units (IMUs), connected to the biped.
Down here, some of the more advanced concepts of LGP have been used (or will be used soon, not yet fully implementad) in the evolution. That is, conditional branching, full operator set, and modularizations (ADFs, subroutines). Also, additional sensor inputs will be included soon.
Preliminary Experiments These preliminary experiments was carried out in the early stages of the project, between october 2002 and january 2003. These results where published in the BizSim symposium of the ASTC'03 conference, March 30 - April 3 2003 in Orlando, Florida, USA. See ref.[2]. Animations: Publications [1] Wolff, K., and Nordin, P. (2003). "Learning Biped Locomotion from First Principles on a Simulated Humanoid Robot using Linear Genetic Programming." In Proceedings of the Genetic and Evolutionary Computation Conference, GECCO 2003 (pp.). Chicago, 12-16 July 2003. Morgan Kaufmann. [2] Wolff, K., and Nordin, P. (2003). "Evolutionary Learning from First Principles of Biped Walking on a Simulated Humanoid Robot." In Proceedings of The Advanced Simulation Technologies Conference 2003, ASTC'03. March 30 - April 3. Orlando, Fl, USA . These papers can be downloaded from Krister Wolff's publications page.
|