BOLeRo - Behavior Optimization and Learning for Robots¶

BOLeRo makes behavior learning for robots easy. It combines behavior learning algorithms with learning problems through defined interfaces that can be used by a controller. The whole communication between a behavior and an environment is done by requesting actions generated from the behavior and sensory information generated by an environment.

The highlights of BOLeRo are

• Dynamical Movement Primitives with a final velocity greater than 0 (see Muelling et al.) and correct handling of rotation in Cartesian space (see Ude et al.) in C++ with a Python wrapper

• policy search:

  • Relative Entropy Policy Search (REPS, see Peters et al.)
  • a clean and readable implementation of Covariance Matrix Adaption Evolution Strategies (CMA-ES, see Wikipedia) and several of its variants

• contextual policy search:

  • Contextual Relative Entropy Policy Search (see Kupcsik et al.)
  • Contextual Covariance Matrix Adaptation Evolution Strategies (see Abdolmaleki et al.)

• it is easy to combine it with the simulation software MARS
• C++ interface to Python modules and Python interface to C++ modules
• configuration of experiments via YAML