Active inference for robot manipulators. Simulation using Franka Emika Panda 7-DOF

How to cite this work

This simulation was developed to support the theoretical results of this work:

• Pezzato C., Ferrari, R., Henrandez C., A Novel Adaptive Controller for Robot Manipulators Based on Active Inference, IEEE Robotics and Automation Letters, 2020.

• @ARTICLE{Pezzato2019, author={C. {Pezzato} and R. {Ferrari} and C. H. {Corbato}}, journal={IEEE Robotics and Automation Letters}, title={A Novel Adaptive Controller for Robot Manipulators Based on Active Inference}, year={2020}, volume={5}, number={2}, pages={2973-2980},}

If you found this controller useful, please consider citing the paper above.

News

Video

A video showing the capability of active inference in the real world is now on youtube!

https://www.youtube.com/watch?v=Vsb0MzOp_TY

Active inference for the real setup

The code for controlling the real Panda using either libfranka or franka_ros is now available at https://github.com/cpezzato/active_inference

Description package for simulation

Repository for the dynamic simulation of the Franka Emika Panda robot arm in Gazebo using Active Inference. This repository contains 3 packages:

• franka_description: it contains the urdf and xacro of the robot and the scene
• panda_control: the package contains the active inference controller and a model reference adaptive controller for comparison
• panda_simulation: package for launching the simulation loading a jointEffortController for each joint and the scene with the robot(s)

The file panda_arm.xacro has been updated with more accurate inertia tensor and centers of mass obtained using Meshlab and the .stl models. See also https://github.com/erdalpekel/panda_simulation

Installation

Make sure you installed ros-kinetic-desktop-full and that you have all the necessary dependencies.

How build a container with the necessary dependancies

If you want, you can install Singularity and build a container for this simulation. You can do that building a singularity image from this singulatity recipe:

https://github.com/cpezzato/panda-xenial-docker/blob/master/panda-recipe-minimal-xenial

Run a shell within the container:

• With Nvidia GPU: $ singularity shell --nv YOUR_IMAGE_NAME
• Without Nvidia GPU: $ singularity shell YOUR_IMAGE_NAME

Once inside the image:

• Source: $ source /opt/ros/kinetic/setup.bash

How to run the simulation

Either in your local machine or inside the image you just created:

• Create a folder for your catkin_ws: $ mkdir -p your_catkin_ws/
  
• Move to the folder: $ cd your_catkin_ws/`
  
• Clone the repository $ git clone https://github.com/cpezzato/panda_simulation.git src
  
• Build the workspace: $ catkin_make
  
• Source: $ source devel/setup.bash
  
• Launch: $ roslaunch panda_simulation simulation.launch

The launch file launches a Gazebo simulation in pause with a single robot. The launch file run the node panda_control_AIC for the active inference controller (AIC) by default. Alternatively one can run the model reference adaptive controller (MRAC) through the node panda_control_MRAC, modifying the launch file accordingly. Just press play and enjoy.

The result with the active inference controller is the following:

Troubleshooting

If you are using ROS Melodic and Gazebo 9, the robot meshes will probably not be showed. To solve this, just remove the "transparent" attribute in the .dae flesin franka_description/meshes/visual