Learning Local Recoveries for Navigation from Human Demonstrations

Source code of the Learning-by-Demonstration framework illustrated in

Del Duchetto, F., Kucukyilmaz, A., Iocchi, L., & Hanheide, M. (2018). Do Not Make the Same Mistakes Again and Again: Learning Local Recovery Policies for Navigation From Human Demonstrations. IEEE Robotics and Automation Letters, 3(4), 4084-4091.

The code is based on ROS kinetic and it runs the experiments using the MORSE simulator.

Installation

1. Install TMuLE: pip install tmule.

2. Install GPy: pip install GPy.

3. Install the STRANDS repository (quick setup in https://strands.readthedocs.io/en/latest/quick_setup.html). In particular you will need the following packages strands_desktop, strands_movebase, strands_morse, topological_navigation, mongodb_store.

4. Clone the Petri Net plans repository (costmap branch) from https://github.com/francescodelduchetto/PetriNetPlans/tree/costmap.

5. Clone this repository.

6. Create a catkin_workspace (http://wiki.ros.org/catkin/Tutorials/create_a_workspace)

7. Link PetriNetPlans/PNPros/ROS_bridge/pnp_ros/pnp_ros, PetriNetPlans/PNPros/ROS_bridge/pnp_ros/pnp_msgs, and learning_local_recovery_navigation into the src/ folder of your catkin workspace, then (when inside your workspace folder) build everything with catkin_make.

Configuration

• Edit the file launch/launch_tmule.launch to put the absolute path to the config folder for the TMULE_CONFIG argument.

• Edit the scripts/setup.sh file to put the path to your catkin workspace.

• Create the empty directory ~/mongodb/test, or in you preferred place but change accordingly MONGO_FOLDER in scripts/setup.sh.

Executing

1. roslaunch llrn launch_tmule.launch

2. Open 0.0.0.0:9999 in a browser.

3. Launch all the windows up to the one named create_topological_map, then create a topological map as explained in link.

4. Stop the window create_topological_map and start all the remaining.

Now you can execute a navigation plan (one of those in pnp_ros/plans, or you can write your own and generate it by pnpgen_linear YOUR_PLAN_NAME.plan YOUR_EXECUTION_RULES.er) by publishing on the topic /planToExec, e.g. rostopic pub /planToExec std_msgs/String "data: 'passdoor12'".

You will be able to signal a failure by pressing the signal button, after some signalling the model should start detecting some similar failure (that you can discard or confirm). After a confirmation you are asked to demonstrate a recovery trajectory that will be automatically used to learn the recovery regression model.

Maintainer: [email protected]