English UT







Joining Lab




Year: 2006-
Atsushi Harada
Kenji Suzuki

- Cognitive Robotics
- Emerging Technologies

Machine Awareness
Action oriented probabilistic self-modeling for a humanoid robot


In recent years, diversification of robot’s workspace has progressed significantly. A robot that uses manipulators with multiple degrees of freedom such as a humanoid robot is expected to operate appropriately in an unknown environment. Manipulators and robot arms with redundant degrees of freedom have serious problems with respect to self-collision and collisions with the external environment, both of which may cause trouble for the robot. The self-body detection and collision detection are often carried out with predetermined rules based on the body structure and knowledge of surrounding environment. The path planning is then performed using collision avoidance technique in order to guarantee a safe operation within given operating ranges.

We proposed a novel method of probabilistic self-modeling based on learning of operating space from exploratory actions of a humanoid robot. Considering a measure to ensure self-preservation in nature, the anthropomorphic humanoid robot learns both of the operating range in each joint and the probabilistic operating space based on Gaussian Mixture Model and Variational Bayesian learning algorithm. The operating space is acquired by using the history of irregular overload, which is detected by using analogue current signals measured by solely internal sensor of joint motors. In addition, online behavior learning with a simple probabilistic path planning is also introduced based on the obtained probabilistic operating space. We will conduct several experiments with a real humanoid robot arm. After the basic characteristics of the obtained operating space are shown, the performance of interaction with different situations such as different load given to the arm and obstacles placed in the surrounding environment will also be demonstrated.

The advantage of the proposed model is scalability because it can be obtained from internal sensors and also extended to integrate with other sensors and predefined model of surrounding environment. Utilizing a probabilistic model to learn the robot’s own behavior provides: system robustness, adapting to the dynamical environment and the acquisition of body representation. For instance, the proposed method allows the robot to track and reach an object in the acquired operating space while it has been updated at the same time.


This work is partly supported by Grants-in-Aid for Scientific Research, MEXT, Japan.

This study was supported in part by the Global COE Program on "Cybernics: fusion of human, machine, and information systems.”

This is a collaboration work with Scuola Superiore Sant'Anna, Italy.

  • Harada, A., Suzuki, K., "Action oriented bayesian learning of the operating space for a humanoid robot, Proc. of IEEE Intl. Conf. on Robotics and Biomimetics (ROBIO), pp. 633-638, 2009
  • Harada, A. and Suzuki, K., "Action oriented self-modeling and Motion Planning for a humanoid robot," Proc. of 2008 IEEE/RAS Intl. Conf. on Humanoid Robots (Humanoids 08), Korea, pp. 367-372, 2008.
  • Harada, A. and Suzuki, K., "Active acquisition of operating ranges and path planning for a humanoid robot," Proc. of IEEE Intl Conference on Robotics and Biomimetics, pp.739-744, China, 2007
Related Projects


  © 2005-2011 Artificial Intelligent Laboratory, University of Tsukuba, Japan