policy_learning: policy_improvement | policy_improvement_loop | policy_improvement_loop_test | policy_learning_tools | policy_library | policy_msgs | task_manager

Stack Summary

policy_learning

• Author: Maintained by Mrinal Kalakrishnan, Peter Pastor
• License: BSD
• Repository: usc-clmc-ros-pkg
• Source: git https://github.com/usc-clmc/usc-clmc-ros-pkg.git

New github repo: [https://github.com/usc-clmc/usc-clmc-ros-pkg]

Documentation

This stack contains code that implements trajectory-based reinforcement learning. Policies are contained in the policy_library package, currently contains DMPPolicy (Dynamic Movement Primitives), and Covariant Trajectory Policy (discretized trajectories). The PI^2 (Policy Improvement with Path Integrals, Theodorou et al, 2010) algorithm is implemented in the policy_improvement package.

A "Task" is a motion that the robot can perform, along with an associated cost function. Once a task is implemented, the system will perform the task repeatedly and optimize the cost function. New "Task"s can be implemented by deriving from the "Task" abstract class in task_manager. Examples tasks are contained in "pr2_tasks".

Video