Humanoids2005    IEEE-RAS International Conference on Humanoid Robots

 
Invited Talk


1) Neuroscience and Humanoid Robots - Hierarchiscal, Modular
Reinforcement Learning with MOSAIC

Dr. Mitsuo Kawato
ATR Fellow, and Director of
ATR Computational Neuroscience Labs
http://www.cns.atr.jp/%7Ekawato/

Date: Dec.6th, 9:10-10:10
Place: Room 101



Computational neuroscience is the discipline which integrates
 humanoid robotics and traditional empirical studies of brain by
 proposing computational theories, implementing them on humanoids for
 examination of their computational competence, and finally examining
 them in carefully designed biological experiments.  In ATR
 Computational Neuroscience Laboratories, several theories have been
 proposed.  For example, cerebellar internal models, MOSAIC,
 basal-ganglia Q learning, neuromodulator hypothesis and so on.  All
 of these were recently supported by experiments and some have been
 implemented on humanoid robots.  On robotics side, supervised
learning of internal models, imitation learning, reinforcement
 learning, three different eye movements, 3-D object recognition with
 tracking, biologically motivated biped locomotion have been
 successfully implemented on several humanoid robots including DB, CB
 and Qrio.  My own recent theoretical interests are around
 hierarchical, modular reinforcement learning scheme with MOSAIC
 (MOdular Selection And Identification Controller) model.  Here, in
 the lower level with continuous-time and continuous-state space, the
 continuous reinforcement learning problem is solved while
 discretizing the state space with a finite number of predictors.  On
 the top level, discrete-time discrete-space Q-learning problem is
 solved with the discretization by the lower level.  The upper level
 actions define low-level reward functions (sub-goals) and ultimately
 determine value functions.  Sub-goals are updated based on the real
 reward from the environment.  Thus, the proposed scheme is the first
 framework of the hierarchical reinforcement learning which does not
 depend on ad-hoc manually tuned divisions of the state space or
 sub-goals.

2) Humanoid Robots Step Up to Perform Human Tasks - Our R&D Challenge

Mr. Tadahiro Kawada,
President of Kawada Industries Inc
.


Date: Dec.7th, 9:00-10:00
Place: Room 101






Kawada Industries Inc. was launched on the development of humanoid robots
when it was involved in H6 and H7 projects of the University of Tokyo.
This talk presents our technological background leading to the projects at first. 
Kawada Industries Inc. then attended the Humanoid Robotics Project of METI,
the Ministry of Economy, Trade and Industries of the Japanese Government,
and had developed HRP-2 which is the first human-size humanoid robot that can get up and lie down as well as fall down safely.  Kawada Industries Inc. had developed a biped walking dinosaur robot that has been displayed in Expo 2005 Aichi.  This talk overviews the history of the developments.  The current activities of Kawada Industries Inc. are also presented with a future plan and perspectives.


Go to Session Plan