Elsevier

Neural Networks

Volume 23, Issue 3, April 2010, Pages 452-460
Neural Networks

A hybrid CPG–ZMP control system for stable walking of a simulated flexible spine humanoid robot

https://doi.org/10.1016/j.neunet.2009.11.003Get rights and content

Abstract

Biped humanoid robots have gained much popularity in recent years. These robots are mainly controlled by two major control methods, the biologically-inspired approach based on Central Pattern Generator (CPG) and the engineering-oriented approach based on Zero Moment Point (ZMP). Given that flexibility in the body torso is required in some human activities, we believe that it is beneficial for the next generation of humanoid robots to have a flexible spine as humans do. In order to cope with the increased complexity in controlling this type of robot, a new kind of control system is necessary. Currently, there is no controller that allows a flexible spine humanoid robot to maintain stability in real-time while walking with dynamic spine motions. This paper presents a new hybrid CPG–ZMP control system for the walking of a realistically simulated flexible spine humanoid robot. Experimental results showed that using our control method, the robot is able to adapt its spine motions in real-time to allow stable walking. Our control system could be used for the control of the next generation humanoid robots.

Introduction

Over the past few years, there has been an increasing interest in studying biped locomotion within the robotics community. Traditionally, two main approaches have been used to control biped walking robots. The most commonly used one is based on the ZMP criterion proposed by Vukobratovic, Frank, and Juricic (1970). The ZMP is a point on the ground where the moments (due to gravity and body inertia) about the two axes forming the ground are zero. The core idea of the ZMP criterion is that if the ZMP of a biped robot is within the support polygon made between the foot and the ground, then stability is guaranteed (Vukobratovic et al., 2001, Vukobratovic and Borovac, 2004, Vukobratovic et al., 1990). By ensuring that the ZMP is within the support polygon during the motion planning stage, researchers are able to make their robots walk stably. Although this approach has been successfully applied to control robots such as the ASIMO, QRIO and the WABIAN-2 (Nagasaka et al., 2004, Ogura et al., 2006, Sakagami et al., 2002), it has several drawbacks. For instance, it required an accurate model of the robot and its walking environment (Endo et al., 2004, Ijspeert, 2003, Kajita et al., 2003). Also, to generate dynamic walking gaits, expensive high performance actuators are desirable. Originally, the ZMP-based approach required that the robot motions be planned off-line ahead of time (Takanishi, 1993, Takanishi, 1988). Recently, researchers have proposed more robust control methods to allow robots to cope with perturbations in real-time (Harada et al., 2004, Hirai et al., 1988, Huang et al., 2001, Hyon et al., 2007, Nagasaka et al., 2004, Park and Chung, 1999, Sugihara and Nakamura, 2005, Sugihara et al., 2002).

The second approach is based on the concept of CPG commonly found in neuroscience literature (Calancie et al., 1994, Capaday, 2002, Choi and Bastian, 2007, Cohen, 1988, Delcomyn, 1998, Ekeberg and Pearson, 2005, Getting, 1986, Getting, 1988, Grillner, 1985, Grillner, 1996, Grillner et al., 1988, Grillner et al., 1991, Ikemoto and Yu, 2008, Kandel et al., 2000, Orlovsky et al., 1999, Tamburin et al., 2007, Zehr et al., 2007). It is known that within the spinal cord of vertebrates, there are networks of neurons responsible for generating rhythmic motions for animals. Controllers based on this approach are more adaptive than those based on the ZMP approach. Furthermore, motions can be generated in real-time with less computation. The problem with this approach is that it is difficult to set up parameters for the neural networks to coordinate all the actuators of a high degrees of freedom walking robot with a flexible spine. Examples of real robots controlled by this approach are the Fujitsu’s HOAP 1 and 2 humanoid robots (Nagashima, 2003, Shan and Nagashima, 2002). (For CPG control of leg joints for simulated walking humanoids, refer to (Hu et al., 1998, Katayama et al., 1995, Miyakoshi, 1998, Taga, 1994, Taga, 1991).)

Note that thus far, researchers working on biped locomotion either assume that their models consist of two legs only or assume that the torso of their models can be represented by a large mass or a stick (Katayama et al., 1995, Park and Rhee, 1998). Although these assumptions simplify the research problem, it is less realistic because humans walk and exhibit behaviors with spine motions. We use our spines in daily activities without consciously thinking. For instance, we bend our body to enter the car, twist our torso to follow an interesting target, avoid a flying object, and express emotions using body postures, etc. Given the importance of the spine in various human activities, it would be beneficial for humanoid robots in the future to have a flexible spine.

In order for the high degree of freedom, flexible spine humanoid robots to perform useful tasks, they need to be able to move from one place to another. Moreover, they should be able to adapt to the environment in real-time. By combining the CPG and ZMP approaches, we seek to create a new type of hybrid control system for the next generation humanoid robots. As a first step, we developed a hybrid CPG–ZMP controller to control a simple belly dancing robot (Or & Takanishi, 2004). The robot consisted of four segments with one-DOF joint in between neighboring segments. At the bottom of the spine, we placed a plastic plate, which served as the footsole. On top of it, we mounted a tilt sensor to measure the tilting angle of the footsole. The CPG-module generated rhythmic body undulations for the spine independently, without knowing the physical specifications of the robot. In other words, no internal model of the robot was required. While the robot was exhibiting belly dance like motions, the signals from the tile sensor was fed back to the control PC through an A/D converter. As the robot was on the verge of falling, a large torque was generated at the base joint. When the torque exceeded an experimentally predetermined threshold, a negative feedback signal was sent to the CPG-module to slow down the spine motions. Experimental results showed that using this hybrid approach, the robot was able to maintain stability in real-time while performing body undulations often seen in belly dancing. Unlike the traditional ZMP-based method, no internal model of the robot was needed. The resulting robot behavior emerged from dynamic interactions between the robot, its controller and the environment.

As the ability to adapt to the environment in real-time is a very challenging task for humanoid robots, in this paper we present a new hybrid CPG–ZMP control system which allows a realistically simulated flexible spine humanoid robot to walk stably with actuated spine motions.

Section snippets

Background

In this section we give an overview of the CPG and ZMP modules of our control system. We then present our humanoid robot model and the Webots simulation software. Finally, we introduce our hybrid controller.

Experiments

To determine the sensitivity of our controller to inputs from the brainstem, we tested it under global excitation ranges from 0.2 to 1.0 (in steps of 0.1) and extra excitation ranges from 0% to 200% (in steps of 100%). Recall that global excitation is the excitation which the brainstem applies to all the neurons in the CPG, and extra excitation is the excitation applied only to the neurons located in the five segments closest to the head. The extra excitation is a percentage of the global

Discussion

In order for the next generation of humanoid robots to behave and interact with humans more naturally, robot designers should pay more attention to the contributions of the spine. In seeking to imitate the flexibility of the human spine commonly observed during locomotion, this research has investigated the possibility of developing a realistically simulated flexible spine humanoid robot that walks stably with dynamic spine motions. As part of our investigation, we propose a new hybrid CPG–ZMP

Conclusion

The current generation of humanoid robots is able to walk stably. However, their torsos act like large rigid body trunks. We believe that it is beneficial for the next generation of humanoid robots to have a flexible spine. The ability to maintain stability in real-time is thus very important for these robots to adapt to the environment. This paper investigates the possibility of using a combination of biologically-inspired and engineering-oriented approaches to control a realistically

Acknowledgement

The author would like to thank Atsuo Takanishi for his support.

References (53)

  • Ö. Ekeberg et al.

    Computer simulation of stepping in the hind legs of the cat: An examination of mechanisms regulating the stance-to-swing transition

    Journal of Neurophysiology

    (2005)
  • Endo, G., Morimoto, J., Nakanishi, J., & Cheng, G. (2004). An empirical exploration of a neural oscillator for biped...
  • P.A. Getting

    Understanding central pattern generators: insights gained from the study of invertebrate systems

  • P.A. Getting

    Comparative analysis of invertebrate central pattern generators

  • S. Grillner

    Neural control of vetebrate locomotion–central mechanisms and reflex interaction with special reference to the cat

  • S. Grillner

    Neural networks for vertebrate locomotion

    Scientific American

    (1996)
  • S. Grillner et al.

    Neural control of locomotion in lower vertebrates

  • S. Grillner et al.

    The lamprey locomotor system — Basic locomotor synergy

  • Harada, K., Kajita, S., Kaneko, K., & Hirukawa, H. (2004). An analytical method on real-time gait planning for a...
  • Hirai, K., Hirose, M., Haikawa, Y., & Takenake, T. (1998). The development of Honda humanoid robot. In Proceedings of...
  • Hu, J., Pratt, J., & Pratt, G. (1998). Stable adaptive control of a bipedal walking robot with CMAC neural networks. In...
  • Q. Huang et al.

    Planning walking patterns for a biped robot

    IEEE Transactions on Robotics and Automation

    (2001)
  • S. Hyon et al.

    Full-body compliant human-humanoid interaction: Balancing in the presence of unknown external force

    IEEE Transactions on Robotics

    (2007)
  • Ijspeert, A. J. (2003). Biologically inspired artificial intelligence....
  • Y. Ikemoto et al.

    The roles of cpg phase modulation and reflexive muscular patterns in balance recovery during walking - a simulation study

    Applied Bionics and Biomechanics

    (2008)
  • Kajita, S., Kanehiro, F., Kaneko, K., Fujiwara, K., Harada, K., & Yokoi, K. et al. (2003). Biped walking pattern...
  • Cited by (40)

    • Robust biped locomotion using deep reinforcement learning on top of an analytical control approach

      2021, Robotics and Autonomous Systems
      Citation Excerpt :

      During training, the framework tries to learn how to generate the walking trajectories, subject to an objective function. In the fourth point of view, the framework is designed by combining the aforementioned approaches [19,22–25]. This type of framework is generally known as a hybrid walking framework.

    • Particle Swarm Optimization aided PID gait controller design for a humanoid robot

      2021, ISA Transactions
      Citation Excerpt :

      The fuzzy PID controller tracks the movement trajectory of the different robot manipulators and finally analyzes the efficiency of the end effectors and locomotion of the robot. Or [26] has developed a new hybrid (central pattern generator) CPG-ZMP control method for robot dynamics, which allows the balancing of a humanoid robot in the real time frame with a flexible spinal system. The flexible spinal system helps in efficient traversing over rough terrains and greatly improves the stability of the humanoid robot.

    • Adaptive Natural Oscillator to exploit natural dynamics for energy efficiency

      2017, Robotics and Autonomous Systems
      Citation Excerpt :

      This capacity is exploited in Central Pattern Generator (CPG) Networks where frequency, amplitudes, and coupling-strengths of a group of oscillators parameterize the motion-generation; see [23] for a survey on CPG-networks and their applications. The provided parameter space is often explored to satisfy objective functions such as stability, locomotion speed, and cost of transportation; for example, in [24] ZMP constraints are satisfied by proper selection of CPG parameters. This exploration – i.e. parameter search – has been carried out using different techniques such as optimal control [25], reinforcement learning [26], and genetic algorithms [27].

    • Online bio-inspired trajectory generation of seven-link biped robot based on T-S fuzzy system

      2014, Applied Soft Computing Journal
      Citation Excerpt :

      Taga [11,12] generate robust biped locomotion using CPG with appropriate feedback signals in an unpredictable environment. Or [13] presents a new hybrid CPG–ZMP control system that is inspired from flexibility in the human body torso. He simulates the flexible spine of a humanoid robot and experimentally shows that the robot can adapt its spine motions in real-time to achieve stable walking.

    View all citing articles on Scopus
    View full text