Projects

Phase 2 Projects

Autonomous Learning of Bipedal Walking Stabilization


Project leaders: Sven Behnke (Bonn)

Researchers: Marcell Missura (Bonn)

Administration: Marion Romagna, Nicola Kokisch (Bonn)

Associates:

Summary:

The first steps of a human child are no more stable than the walking motions of most bipedal robots. But unlike robots, humans quickly learn how to walk in a stable manner. Without much need of supervision, they develop the skills to walk on virtually any surface and to recover from strong, unexpected disturbances. Bipedal walkers are complex, inherently unstable systems. Because of this complexity, the efficiency, grace, and, in particular, the stability of the natural human gait has resisted decades of scientific effort to synthesize a bipedal gait with human-like capabilities.
Starting off with a semi-stable open-loop gait engine and the ability to fall, our new  approach is to develop autonomously learning algorithms that enable a humanoid robot to discover its own body dynamics during disturbed walking. The disturbances will be external or self-generated using an efficient exploration strategy. The robot will use its knowledge to decide when and where to place capture steps in order to recover from unexpected pushes and to cope with uneven surfaces. Our learning method is based on generic non-parametric regression rather than on explicit modeling.

Phase 1 Projects

Autonomous Learning of Bipedal Walking Stabilization


Project leaders: Sven Behnke (Bonn)

Researchers: Marcell Missura (Bonn)

Administration: Marion Romagna, Nicola Kokisch (Bonn)

Associates:

Summary:

The first steps of a human child are no more stable than the walking motions of most bipedal robots. But unlike robots, humans quickly learn how to walk in a stable manner. Without much need of supervision, they develop the skills to walk on virtually any surface and to recover from strong, unexpected disturbances. Bipedal walkers are complex, inherently unstable systems. Because of this complexity, the efficiency, grace, and, in particular, the stability of the natural human gait has resisted decades of scientific effort to synthesize a bipedal gait with human-like capabilities.
Starting off with a semi-stable open-loop gait engine and the ability to fall, our new  approach is to develop autonomously learning algorithms that enable a humanoid robot to discover its own body dynamics during disturbed walking. The disturbances will be external or self-generated using an efficient exploration strategy. The robot will use its knowledge to decide when and where to place capture steps in order to recover from unexpected pushes and to cope with uneven surfaces. Our learning method is based on generic non-parametric regression rather than on explicit modeling.