LEARNING APPROACH TO STUDY EFFECT OF FLEXIBLE SPINE ON RUNNING BEHAVIOR OF A QUADRUPED ROBOT

Effect of flexibility in the spine of a quadruped robot on its energy efficiency and stability is studied. The approach is based on learning parameters of cyclic position command to the actuators. The learning system employs a reinforcement learning method to find the parameters that result in stable running while the velocity-energy ratio is minimized. The learning is done on a simulated system and the final results are tested on a situated robot. The results show that a wider range of parameters results in stable running for softer spine. In addition, robot with harder spine consumes more energy.

[1]  Marc H. Raibert,et al.  Walking BigDog: Insights and challenges from legged robotics , 2006 .

[2]  Fumiya Iida,et al.  Self-Stabilization and Behavioral Diversity of Embodied Adaptive Locomotion , 2004, Embodied Artificial Intelligence.

[3]  Julianna M. Gál.  Spinal flexion and locomotor energetics in kangaroo, monkey, and tiger , 1992 .

[4]  Alex M. Andrew,et al.  Reinforcement Learning: : An Introduction , 1998 .

[5]  J. Gál,et al.  Mammalian spinal biomechanics. I. Static and dynamic mechanical properties of intact intervertebral joints. , 1993, The Journal of experimental biology.

[6]  Martin Buehler,et al.  Quadruped trotting with passive knees: design, control, and experiments , 2000, Proceedings 2000 ICRA. Millennium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No.00CH37065).