Stochastic Optimization of Bipedal Walking using Gyro Feedback and Phase
Resetting
- Authors: Felix Faber and Sven Behnke
- In Proceedings of: IEEE-RAS 7th International Conference on Humanoid Robots (Humanoids), Pittsburgh, USA,
12/2007.
- Abstract:
We present a method to optimize the walking pattern of a humanoid robot
for forward speed using suitable metaheuristics. Our starting point is
a hand-tuned open-loop gait that we enhance with two feedback control
mechanisms. First,
we employ a P-controller that regulates the foot angle in order to
reduce angular velocity of the robot’s body. Second, we introduce
a phase resetting mechanism that starts the next step at the moment of
foot contact. Using a physics-based simulation, we demonstrate that
such feedback control is essential for achieving
fast and robust locomotion.
For the optimization of open-loop parameters and parameters of the
feedback mechanisms, we compare Policy Gradient Reinforcement Learning
(PGRL) and Particle Swarm Optimization (PSO). To make optimization more
data-efficient, we extend PGRL by an adaptive step size and a
sequential sampling procedure. Our experiments show that the proposed
extensions increase the performance of PGRL significantly.
We selected the extended PGRL algorithm to optimize the gait of a real
robot. After optimization, the robot is able to walk significantly
faster..
back
to selected robotic soccer publications
back
to selected machine learning publications