Loading...

Embodiment Approach Used by a Quadruped Robot to Learn How to Walk

Davari, Hossein | 2015

495 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 48228 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Bagheri Shouraki, Saeed
  7. Abstract:
  8. The researches done on legged robots can be categorized in two main groups: based on analytical methods and intelligent methods. All researched done by intelligent methods are similar in adaptation of learning methods. Robots can improve their actions with learning and there is no need to do an exact modeling of environment and system. Also this is adaptive to changes in system and environment. There are different approaches for studying intelligence in cognitive science. Embodiment approach focuses on body, its properties and connection to environment. Controller, morphology and environment effect doing the tasks in embodiment. Morphology contains sensors, shapes of elements, actuators and material. Complicated tasks could be done in this way in a simple and cheap manner. Also expertise is not concentrated in central controller. In this research embodiment approach and EMBER principle are used for a task: learning to walk by a quadrupedal robot. The literature review shows that only two works focus on this topic. In these works the main drawback is that convenient gaits and paths are not obtained for legs. Using them is so important and makes this research closer to embodiment approach in reality. For example children move their legs and learn a convenient locus for their body before they can even stand. The main procedure used for this research can be divided into two parts. First, obtaining the locus for each leg to minimize the consumed energy and maximize the advance. In order to achieve this, particle swarm optimization method (PSO) is used to find the arbitrary and elliptical loci. The objective function is calculated with WEBOTS and returned to the PSO algorithm. In order to determine the joints’ angles a numerical inverse kinematic method, CCD is adapted. The analysis of the whole body movement and determination of optimized locus is very complicated. This difficulty is overcome by using EMBER principle. Determination of the arbitrary locus is a novel method in path planning with PSO. The results of arbitrary locus is much more beneficent than the elliptical locus. The second main part is about cooperation learning of the four legs. This is most important parameter in stability and walking speed. In most of the published literature, a proportional chosen gait is assumed and other parameters are determined by the learning algorithm. The arbitrary locus is used for each leg and the harmony between legs is obtained by Q-Learning algorithm. Similar to previous part, reward function is calculated with WEBOTS and decision matrix is updated. The objective here is to minimize the time and deviation from straight line and also to maximize the advance. Deviation from straight line is quantified by two methods: displacement of the center of gravity relative to straight line and the rotation of robot. Using the second method results in a more straight walking
  9. Keywords:
  10. Machine Learning ; Optimization ; Gait Cycle ; Cognitive Science ; Quadruped Robot ; Embodiment Approach ; Reinforcement Learning

 Digital Object List

 Bookmark

No TOC