Minoru Asada <asada@robotics.ccm.eng.osaka-u.ac.jp> heads a robotics laboratory, the Asada Lab, under the Group of Mechanical Engineering at Osaka University in Japan. Their publications include both low-level robot sensing and control work and higher-level behavior learning.
| Jay : game learning : soccer : Asada Lab |
They use both simulations and real robots. The simulator is good enough that they can transfer learned behavior directly from a simulated robot to a real one, saving a lot of time and hardware expense. The robots are larger than the Dynamites. They have tracks, and each one carries a video camera.
Don't be daunted by the number of papers. Being self-contained, the papers have a lot of boilerplate, and each one presents only a modest amount of new information.
Vision-based behavior acquisition for a shooting robot by using a
reinforcement learning (1994, 7 pages)
Minoru Asada,
Shoichi Noda, Sukoya Tawaratsumida and
Koh Hosoda
Compressed postscript. Learning to shoot the ball into the goal,
without any opponent or teammate, with Q-learning.
All input information comes from a camera image.
A vision-based reinforcement learning for coordination of
soccer playing behaviors (1994, 6 pages)
Minoru Asada,
Eiji Uchibe,
Shoichi Noda, Sukoya Tawaratsumida and
Koh Hosoda
Compressed postscript.
Combines the goal-shooting behavior of the previous paper with
a separately-learned behavior to avoid moving obstacles,
so the robot can shoot without running into things, and can
shoot past a (sufficiently stupid) goalie.
Purposive behavior acquisition on a real robot by
vision-based reinforcement learning (1994, 9 pages)
Minoru Asada, Shoichi Noda, Sukoya Tawaratsumida and
Koh Hosoda
Compressed postscript. Little new information.
Focuses on problems in getting the robot to perform well, and
their solutions.
Coordination of multiple behaviors acquired by
a vision-based reinforcement learning (1994, 8 pages)
Minoru Asada,
Eiji Uchibe,
Shoichi Noda, Sukoya Tawaratsumida and
Koh Hosoda
Compressed postscript.
Condiders three ways to combine robot behaviors which are dependent
on each other, such as the behaviors of avoiding an opponent and
shooting for a goal.
Vision-based reinforcement learning for purposive behavior acquisition
(1995, 8 pages)
Minoru Asada,
Shoichi Noda, Sukoya Tawaratsumida and
Koh Hosoda
Compressed postscript. Little new information.
Non-physical intervention in robot learning based on LfE method
(1995, 7 pages)
Minoru Asada, Shoichi Noda and
Koh Hosoda
Compressed postscript.
"Intervention" means giving the robot extra information to help it learn.
"LfE" means learning from examples.
The paper is actually about getting the robot to construct its own
state- and action-spaces for Q-learning (in earlier papers, the spaces
were provided by a programmer).
Agents that learn from other competitive agents (1995, 7 pages)
Minoru Asada,
Eiji Uchibe and
Koh Hosoda
Compressed postscript.
Also available
here from the
Agents that Learn from Other Agents workshop.
Learning to shoot past a goalie by watching the goalie's behavior.
Vision-based robot learning for behavior acquisition (1995, 6 pages)
Minoru Asada, Takayuki Nakamura and
Koh Hosoda
Compressed postscript.
The robot decides what to do based on its camera images without
bothering to construct a 3D model of its situation.