Learning to Balance Upright Posture: What can be Learnt Using Adaptive NN Models?

  • N. Alberto Borghese
Conference paper
Part of the Lecture Notes in Computer Science book series (LNCS, volume 2486)


Human upright vertical position is unstable. A continuous activation of postural muscles is therefore required to avoid falling down. The problem is complicated by the reduced dimension of the support base (the feet) and by the articulated structure of the human skeleton. Nevertheless, upright posture is a capability, which is learnt in the first year of life. Here, the learning process is investigated by using neural networks models and the reinforcement learning paradigm. After creating a mechanically realistic digital human body, a parametric feed-back controller is defined. It outputs a set of joint torques as a function of orientation and rotational velocity of the body segments. The controller does not have any information either on the model of the human body or on the suitable set of its parameters. It learns the parameters which allow the controller to maintain the vertical position, through trial-and-error (success-fail) by using reinforcement learning. When learning is completed, the kinematics behaviour qualitatively resemble that of real experiments. The amplitude of the oscillations is larger than in the real case; this is due to the lack of any explicit penalization of the oscillations amplitude. The kinematics resembles the real one also when the body has to maintain the vertical upright position facing a tilt or displacement of the support base.


Human posture control 


Unable to display preview. Download preview PDF.

Unable to display preview. Download preview PDF.


  1. [1]
    Pedotti A. A study of motor coordination and neuromuscular activities in human locomotion. Biol. Cybern., 26:53–62, 1977.CrossRefGoogle Scholar
  2. [2]
    Pedotti A., Krishnan V. V., and Stark L. Optimization of muscle-force sequencing in human locomotion. Math. Biosc., 38:57–76, 1978.CrossRefGoogle Scholar
  3. [3]
    A. G. Barto, R. S. Sutton, and C. W. Anderson. Neuronlike adaptive elements that can solve difficult learning problems. IEEE Trans. Syst. Man and Cybern., 13:834–846, 1983.Google Scholar
  4. [4]
    Ghez C. Posture. Elsevier, 1991.Google Scholar
  5. [5]
    de Leva P. Joint center longitudinal positions computed from a selected subset of chandler’s data. J. Biomechanics, 9:1231–1233, 1996.Google Scholar
  6. [6]
    Lacquaniti F., Maioli C., Borghese N. A., and Bianchi L. Posture and movement: Coordination and control. Archives Italiannes de Biologie, 135:353–367, 1997.Google Scholar
  7. [7]
    Massion J. Postural contol system. Curr. Opin. Neurobiol., 4:877–887, 1994.CrossRefGoogle Scholar
  8. [8]
    L. P. Kaelbling, M. L. Littman, and A. W. Moore. Reinforcement learning: A survey. J. Artificial Intelligence Research, 4:237–285, 1996.Google Scholar
  9. [9]
    Kashima, Isurugi Y. T., and Shima M. Analysis of a muscular control system in human movements. Biol. Cybern., 82:”123–131”, 2000.zbMATHCrossRefGoogle Scholar
  10. [10]
    Nashner L.M. Fixed patterns of rapid postural responses among leg muscle during stance. Exp. Brain Res., 30:13–24, 1977.CrossRefGoogle Scholar
  11. [11]
    Nigg B.M. and Herzog W. John Wiley and Sons, Toronto, 1999.Google Scholar
  12. [12]
    Bernstein N. Pergamon Press, Oxford, 1967.Google Scholar
  13. [13]
    Borghese N. A., Bianchi L., and Lacquaniti F. Kinematic determinants of human locomotion. J. of Physiology, 494(3):863–879, 1996.Google Scholar
  14. [14]
    R. Sutton T. Miller and P. Werbos, editors. MIT Press, 1990.Google Scholar

Copyright information

© Springer-Verlag Berlin Heidelberg 2002

Authors and Affiliations

  • N. Alberto Borghese
    • 1
  1. 1.Laboratory of Human Motion Analysis and Virtual Reality (MAVR), Department of Computer ScienceUniversity of MilanoMilano

Personalised recommendations