SongQ-Young / discrete_active_inference

MDP for robotic task planning

discrete_active_inference for robotics

Repository for active inference and behavior trees for discrete decision making. This repository relies on a TIAGo simulation in a simplified retail store. Please read the associated paper for more theorethical considerations about the algorithms.

"Active Inference and Behavior Trees for Reactive Action Planning and Execution in Robotics"

Corrado Pezzato, Carlos Hernandez, Stefan Bonhof, Martijn Wisse, https://arxiv.org/abs/2011.09756

Content

This repositiry contains a Matlab examples and a ROS package for active inference for task planning and execution.

Main files

Matlab:

• aip.m the active inference algorithm for decision making is illustrated in the case of heterogeneous states and actions.
• example.m example of use of active inference for discrete decision making in a robotic case where conflicts and preconditions checks are required. A robot is assumed to be able to navigate to a point (MoveBase), reach a location with its end effector (Move MPC), and pick and place things. Actions have preconditions and are assumed not instantaneous

ROS:

The other folders are related to the ROS package containing a Python implementation of active inference and behavior trees. You can run an example use case with TIAGo in a simplified retail store after installation of the package ad dependancies.

Dependencies

Simulation Environment

A singularity image can be downloaded from here.

Alternatively, you can build the singularity yourself:

1. create a sub directory called 'pkgs' (in the singularity_environment directory)

      mkdir pkgs

2. use vcstool (or wstool) to clone/download the dependencies (as specified in retail_store_lightweight_sim.repos).

      vcs import < retail_store_lightweight_sim.repos pkgs

   Adding packages to pkg will allow rosdep to install all required build and run dependencies into the image, so students can then proceed to build those packages in their own workspaces (otherwise builds would fail due to missing dependencies).

   Note Packages in pkg will be installed on the image, their source will not be included in the image itself, so there may be some elements that are not installed. So far I've only noticed one required change.

3. Modify the CMakeList.txt file from the pal_navigation_sm inside the pkgs folder.

   Change the install instruction (starts at line 10) by adding some scripts as follows.

   install(
   PROGRAMS
      scripts/map_setup.py
      scripts/pal_navigation_main_sm.py
      scripts/navigation.sh
      scripts/base_maps_symlink.sh
      scripts/cp_maps_to_home.sh
      scripts/cp_pose_to_home.sh
      DESTINATION ${CATKIN_PACKAGE_BIN_DESTINATION})

4. check the VERSION variable inside the docker_build.sh, build.sh and Singularity files. This version should match the version of your singularity install (singularity -v)

5. run docker_build.sh

      ./docker_build.sh

   After some time and a successful build, a new docker image will be created. This requires Docker to be installed and configured.

6. run build.sh

      ./build.sh

After some time and a successful build, a new .simg should be generated by singularity in the cwd.

Behavior trees library

Install the BT library to use this package (tested in Ubuntu 18.04 with ROS Melodic). Before proceeding, it is recommended to to install the following dependencies:

sudo apt-get install libzmq3-dev libboost-dev

You can also easily install the Behavior Tree library with the command

sudo apt-get install ros-$ROS_DISTRO-behaviortree-cpp-v3
sudo apt-get update   

Running the code

Using the virtual environment

Access the simngularity image by using the regular Singularity shell action:

singularity shell /path/to/discrete_ai_tiago.simg

Use the flag for nvidia drivers if applicable to your machine:

singularity shell --nv /path/to/discrete_ai_tiago.simg

Then source /opt/ros/melodic/setup.bash to access all the TIAGo dependencies installed on the image.

source /opt/ros/melodic/setup.bash

How to run a simple example with TIAGo

Create a new workspace and clone this repository in the src folder. Build the package using catkin build. Run the three commands below from within the singularity image after sourcing source/devel/setup.bash.

roslaunch retail_store_simulation tiago_simulation.launch
rosrun discrete_ai tiago_perception.py
rosrun discrete_ai active_inference_server.py

From a terminal outside the singularity image run the behavior tree:

rosrun discrete_ai demo_executeBT

The expected outcome is the following:

Note: The sills used in this simulation are based on standard moveBase and moveIt actions, thus robustness (especially of IK solutions) might make TIAGo fail the grasp. Aruco detection can also imprecise and will be improved over time.