Only released in EOL distros:  

Package Summary

rl_msgs is a package of ROS message definitions, which are used for a reinforcement learning agent to communicate with an environment.

Documentation

This package defines a standard way for agents and environments to communicate through ROS messages. In ROS computation is done by nodes that perform different tasks that communicate through a set of ROS messages. ROS provides the framework to publish, subscribe to and receive these messages, in addition to tools for logging, displaying, plotting, and playing back ROS messages. This package defines a set of ROS messages for a reinforcement learning agent and environment to communicate. These messages are similar to the messages used by RL-Glue (Tanner and White 2009), but are simplified and defined in the ROS format.

Please take a look at the tutorial on how to install, compile, and use this package

Check out the code at: https://github.com/toddhester/rl-texplore-ros-pkg

reinforcement_stack_overview.png

Agents from the rl_agent package can communicate with environments from the rl_env package through this interface. In addition, you can write your own agents and environments that will work with the ones in those packages by passing the same messages defined here.

The framework is being designed to be as flexible as possible. While our algorithms and environments will mostly follow a certain style, we understand that every robot and every experiment differs. Our hope is to not force our philosophy onto other people, as long as you adhere to rl_msgs you can design and implement your algorithms and environments in your own style.

Messages passed from Agent to Environment

  • RLAction: This message sends the environment the action (an integer) that the agent has selected. rl_msgs/RLAction

  • RLExperimentInfo: This message provides information on the results of the latest episode of the experiment, namely an episode number, the sum of rewards for that episode, and the number of actions the agent took in that episode. The episode_reward can be plotted to look at rewards per episode. rl_msgs/RLExperimentInfo

Messages passed from Environment to Agent

  • RLEnvDescription: This message describes the environment with a title, the number of actions, the number of states, the range of each state feature, the maximum one-step reward, the range of rewards, if the task is episodic or not, and if the task is deterministic or stochastic. rl_msgs/RLEnvDescription

  • RLEnvSeedExperience: This message provides a full experience <s, a, s', r> seed for the agent to use for learning. In addition to the state, action, next state, and reward, it also includes a boolean indicating if it is a terminal transition or not. rl_msgs/RLEnvSeedExperience

  • RLStateReward: This is a message from the environment with the agent's new state, the reward received on this time step, and whether this transition is terminal or not. rl_msgs/RLStateReward

References

Wiki: rl_msgs (last edited 2015-09-06 01:12:55 by ToddHester)