Hajime Kimura

About

Hajime Kimura is a pioneering researcher in reinforcement learning for robotics, specializing in applying machine learning to high-dimensional, continuous control problems. His major contributions focus on developing algorithms that enable robots to learn complex locomotion and manipulation tasks through trial and error, without explicit programming. Kimura’s most influential work, “Reinforcement learning of walking behavior for a four-legged robot” (2003), with 93 citations, introduced an innovative action selection scheme for actor-critic algorithms, successfully handling an eight-dimensional continuous state/action space to achieve stable walking gaits. He further advanced the field by proposing methods like random tiling and Gibbs sampling to tackle multi-dimensional state-action spaces, as seen in his 2006 paper (6 citations), and by applying natural gradient Actor-Critic algorithms to underwater robots for swimming, walking, and grasping (2009, 8 citations). Kimura’s research bridges theoretical reinforcement learning with real-world robotic applications, demonstrating feasibility on platforms ranging from legged robots to multifunctional underwater systems. His work has been instrumental in making reinforcement learning practical for high-degree-of-freedom robots, inspiring subsequent advances in autonomous motion planning and adaptive control.

Research Focus

Key Achievements

Top Papers

1. 1
   Reinforcement learning of walking behavior for a four-legged robot

   93 citations · 2003
2. 2
   Reinforcement learning of walking behavior for a four-legged robot

   28 citations · 2002
3. 3
   Reinforcement Learning using Stochastic Gradient Algorithm and its Application to Robots

   11 citations · 1999
4. 4
   Unified motion planning method using random network and gradient method for multifunctional underwater robots

   8 citations · 2009
5. 5
   Reinforcement Learning in Multi-dimensional State-action Space Using Random Tiling and Gibbs Sampling

   6 citations · 2006
6. 6
   A Policy Representation Using Weighted Multiple Normal Distribution Real-time Reinforcement Learning Feasible for Varying Optimal Actions.

   4 citations · 2003
7. 7
   Motions obtaining of multi-degree-freedom underwater robot by using reinforcement learning algorithms

   2 citations · 2010
8. 8
   Reinforcement Learning for Crawling Robot Motion Using Stochastic Gradient Ascent

   2 citations · 1999

Key Collaborators

Contact & Links