Overview

Theme

Design and control of anthropomorphic robots

research overview

Motivation

First of all, I'm interested in human intelligence, particularly in the decision making mechanism. In my opinion, the decision making process is mostly ruled by the motor controller embodied. Thus, the controller design is thought to be an important cross-section to approach the problem.

Secondly, I'm interested in controlling complicated dynamical systems like humans. The human motion is generated by a mutual effect of multiple motor actuations, the reaction forces from the world, and the gravity. It is challenging in terms of engineering how to coordinate the behaviors of such systems.

Thirdly, I'm expecting the future utility of the humanoid robot. It is an ultimately personified machine, which possesses the best interface to serve as user's "another body" ad litteram.

Difficulties

Two characteristics lie on humanoid dynamics. One is a large scale of the structure. The humanoid robot has a number of movable parts and motors. Not only the number, but the structual complexity is a problem; trunk, extremities and endpoints don't form a simple hierarchy. The other is the passivity in nature, meaning that the robot is driven only by the reaction force from the external world. It is a common property of mobile robots. But, the causality between the external reaction force and the resultant body movement is complex particularly in the case of humanoids.

On the other hand, three characteristics lie on humanoid systems. One is that it is open to the world, a common difficulty among all robotic systems. Another one is that it is highly integrated. They have to carry all components including a number of motors, sensors, processors, batteries, etc. The last one is the quantity of information to handle. It doesn't only mean that many hardware components have to be controlled, but also that a software framework which users can easily extend for their purposes without bankruptcy is necessary.

Approach

My technical background is in mechano-informatics. In order to model and process human and humanoid motions in both the computer world and the real world, I'm currently proceeding the following researches in aspects of mechanical dynamics, multibody kinetics, automatic control, software engineering, etc.


Multiscale dynamics and control

Multiscale dynamics approach

The equation of motion of humanoid system seems to be complicated. But, an inspection of the equation discloses the following partial relationships within.

The former two relationships represent a macroscopic dynamics of humanoids, while the latter two do microscopic dynamics in the systems. The both are bridged by a mesoscopic dynamics. I design controllers based on such multiscale behaviors.

A separation of bipedal system dynamics into micro-macro controls is a conventional approach(Miyazaki and Arimoto 1980, Furusho et al. 1981, Fujimoto and Kawamura 1995). However, how to bridge them and how to design a controller have been an open problem.

Macroscopic dynamics and control

No matter how a system is dynamically complicated, the reaction force from the external world generates the COM acceleration. It means that a skill of the external force manipulation is important for the control COM behavior (Witt 1968, Yamashita et al. 1972, Fujimoto and Kawamura 1995). A problem is that the external force has natural limits as follows:

The former constraint is more serious than the latter, and is substitutable for the condition where the center of pressure (COP) is inside of the contact area (Vukobratovic et al. 1969). An interesting fact is that the position manipulation of COP is almost equivalent to the reaction force manipulation (Mitobe et al. 1996). A dynamical similarity between COP and COM to a carted inverted pendulum underlies this property(Nagasaka et al. 1999). The problem of this scope is how COP should be manipulated.

I've proposed the following control methods so far.

  1. Dual-term disturbance absorption: Given a referential motion model, it is OK that the robot is controlled so as to imitate the model patterns of COM and COP. When any of them goes away from the model pattern, it will cause a paradox in the control if making both close to the patterns simultaneously. My idea is to make COP follow the pattern in short term, while do COM converge to the pattern in long term.
    [References]
    • Tomomichi Sugihara and Yoshihiko Nakamura, Dual Term Absorption of Disturbance for the Whole-body Cooperative Motion Control of Humanoid Robot (in Japanese), Journal of Robotics Society of Japan, Vol.24, No.1, pp.64-73, 2006.
    • Tomomichi Sugihara and Yoshihiko Nakamura, Whole-body Cooperative Balancing of Humanoid Robot using COG Jacobian, 2002 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp.2575-2580, Lausanne, Oct., 2002.
  2. ZMP manipulation based on carted inverted pendulum: Not given a motion model, it is OK to manipulate COP as if it were the cart of an inverted pendulum. Note that the movable range of COP is limited within the contact region. Although it is hard to achieve sophisticated behavior only by this method, it is flexible against uncertainties and perturbations since it is not slaved by the pre-designed motion trajectory.
    [References]
    • Tomomichi Sugihara and Yoshihiko Nakamura, High Mobility Control of Humanoid Robots Based on an Analogy of ZMP--COG Model and Carted Inverted Pendulum Model (in Japanese), Journal of Robotics Society of Japan, Vol.24, No.1, pp.74-83, 2006.
    • Tomomichi Sugihara, Yoshihiko Nakamura and Hirochika Inoue, Realtime Humanoid Motion Generation through ZMP Manipulation based on Inverted Pendulum Control, 2002 IEEE International Conference on Robotics and Automation, pp.1404-1409, Washington DC, May, 2002.
  3. Simulated regulator: When the COP is stuck at the rim of the contact region, it is OK to step out so as to include the simulated COP position in the contact region in the future. It is also free from the pre-planned motion trajectory.
    [References]
    • Tomomichi Sugihara, Simulated Regulator to Synthesize ZMP Manipulation and Foot Location for Autonomous Control of Biped Robots, 2008 IEEE International Conference on Robotics and Automation, 2008 (to appear).
    • Tomomichi Sugihara, Synthesized COM-Pedipulation Control of Biped Robots Based on Simulated Regulator (in Japanese), 25th Annual Conference of Robotics Society of Japan, Chiba Institute of Technology, Sep., 2007.
  4. Impedance swithcing: By assuming a virtual spring between COM and COP, and switching the virtual impedance, robots can even jump.
    [References]
    • Tomomichi Sugihara and Yoshihiko Nakamura, Contact Phase Invariant Control for Humanoid Robot based on Variable Impedant Inverted Pendulum Model, 2003 IEEE International Conference on Robotics and Automation, pp.51-56, Taipei, Sep., 2003.
    • Tomomichi Sugihara and Yoshihiko Nakamura, Contact Phase Invariant Controller Design for Humanoid Robot based on Inverted Pendulum Model with Variable Impedance (in Japanese), 20th Annual Conference of Robotics Society of Japan, Osaka, Sep., 2002.

Mesoscopic dynamics and control

The problem in this scope is how the external force or COP can be manipulated.

The COP position is directly determined by the ankle motor torque in cases of toeless robots (Sorao et al. 1997, Hirai et al. 1997, Obata et al. 1998, Kajita et al. 1999, Nomura et al. 2002). But, it is not clearly predictable how the upper body will move as the result. According to the dynamical similarity of humanoids to the carted inverted pendulum, the movement of COP is coupled with that of COM (more accurately, the COM acceleration). Hence, the COP manipulation is equivalently achieved if the COM acceleration is manipulated by the actuation of motors. It is a difficult inverse problem to compute the amount of motor actuations from the desired COM acceleration. Though a method to approximate COM motion by the trunk motion is proposed (Nagasaka et al. 1999), it had been a pending issue to improve the accuracy for variety of motions.

By the way, it is a well-known method to compute motor velocities from the desired endpoint velocity in cases of robotic arms (Whitney 1969). As an augmentation of this idea, a method to compute motor velocities from the desired COM velocity has been proposed (Boulic et al. 1995). A problem is how to apply it to legged robots, which are not mechanically connected to the ground, different from robotic arms. Incidentally, a conventional way was the same with that for arms (Tamiya et al. 1997, Hirano et al. 1998).

Against this issue, I proposed an efficient computation method of the COM Jacobian matrix embedding unactuated coordinates in accordance with the classification of contact conditions, and its application to the resolved COM rate control. By combining this Jacobian matrix with other matrices which relates motor velocities to various motion rates such as the trunk rotation velocity and the total angular momentum, the whole-body motions are coordinated. This method is now applied to many humanoid robots in the world.
[References]

Microscopic dynamics and control

The resolved COM rate control works on robust joint servo controllers to compensate effects of the friction and gravity. Many standard techniques have been proposed for robotic arms. In order to make them available on humanoid robots, some arrangements are required.
[References]

Collision and contact with the external world is the largest source of differences on motions from the ideal model. A difficulty to deal with them are particularly put on the forward dynamics computation. A physically stable phenomenon causes numerically unstable manners on computer simulations. I'm working on how to compromise the physical accuracy and numerical stability.
[References]


Motion planning

Motion planning means to search or design a function of time which smoothly connects the start point and the goal point. In many cases, geometric constraints such as collision avoidance are the main issue. But, in the case of legged robots, the above constraints on the external force which will be exerted to the robot have to be additionally satisfied. In this sense, dynamical constraints are also the inseparable issue.

Let us represent the robot motion by COM and COP based on the macroscopic dynamics. Now, the problem is to find a COM trajectory from the start to the goal and also a COP trajectory which satisfies the constraints(Kurematsu et al. 1988, Kajita et al. 1991, Kurazume et al. 2003, Nagasaka et al. 2004, Harada et al. 2005). Since COM and COP mutually affect, it is not easy to handle both conditions simultaneously. Conventional methods either target only slow walking motions, or batch planning of some steps.

The boundary condition relaxation method which I proposed enables a stepwise motion planning by moderating both the COM condition and COP condition. Robots can achieve agile stepping motions including sudden starts, stops and turnings completely in online.
[References]


Robot implementation

Let me introduce the robots of which I have contributed to the developments. Refer to the following pages and publications for more information.

[References]


Motor intelligence

Refer to the following webpage.
Animatronic Humanoid Robot Project at Nakamura and Yamane laboratory in the university of Tokyo
[References]

Top page

Research

Curriculum-Vitae

Publication list