Humanoid push recovery through simple models and learning

Speaker: Ambarish Goswami
Honda Research Institute

Work done by: Jerry Pratt, John Rebula and Ambarish Goswami
IHMC, Florida, and HRI, California

It is known that for a large magnitude push a human or a humanoid
robot must take a step to avoid a fall. Despite some scattered
results, a principled approach towards ``When and where to take a
step" has not yet emerged.  Towards this goal, we use a simple
dynamic model and a learning technique. The simple dynamic model
allows exact analytical solutions, and the learning module
compensates for the model discrepancy.

Specifically, we present the concept of Capture Point, the point
on the ground where a humanoid must step to in order to come
to a complete stop. The location of the Capture Point relative
to the base of support determines which strategy the robot
should adopt to successfully stop in a given situation.

To analytically compute the Capture Point for a humanoid,
we make use of the vastly simplified Linear Inverted Pendulum
Model. While this results in fast computation, model
assumptions and modeling errors lead to stepping in the wrong
place and resulting in large velocity errors after stepping.
To address this we use learning techniques to update the
analytically predicted Capture Points. We validate our method
on a three dimensional humanoid robot simulation with 12
actuated lower body degrees of freedom, distributed mass,
and articulated limbs. Using our learning approach,
robustness to pushes is significantly improved as compared
to using the Linear Inverted Pendulum Model without learning.