- Syed Hasan Shozab Abidi, 4929631@studenti.unige.it
- Muhammad Ali Haider Dar, 5046263@studenti.unige.it
MSc Robotics Engineering, University of Genoa, Italy
Instructor: Prof. Fulvio Mastrogiovanni
This project is an extension of the first assignment of Artificial Intelligence for Robotics II course, in which we got familiar with the operational usage of an AI task planner and tested our understanding by developing a planning model for a robotic coffee shop. In this project, we used the same abstract scenario of the robotic coffee shop in order to get familiarized with the idea of modeling integrated task and motion planning problems in a belief space. The entire project can be found in this repository along with the instructions to compile it.
The context of the assignment is the same, there is a waiter robot which is assigned with a task to serve the orders to the respective customers by visiting their coffee tables in an optimal way. The optimality is based on the distance travelled by the robot. Therefore, the robot has to reason on the discrete sequence of table visits to cover all the regions in the coffee shop.
Unlike the previous assignment, we are considering that uncertainties will be involved with the robot’s motion, thus the robot has to navigate and localize itself while travelling from one region to another. Here ‘region’ refers to the table, in a realistic scenario a robot can be anywhere around the table while serving the customers but for simplicity we are considering that all the regions have a single way-point location associated with them in a 2D grid. It means that if the robot reaches the corresponding waypoint location of the region that region will be considered as being visited.
In real world problems motion planning is done under uncertainties. For example when the robot moves in a real world, there are many sources of uncertainties which can affect its planning such as sensors and motors. The sensors and encoder motors used in robots are not ideal and thus the position and orientation readings the robot retrieves from them are uncertain to some extent. To overcome these uncertainties we incorporate them during the planning phase. In our case we can classify these uncertainties as the robot state uncertainty and we compute it by taking the trace of the covariance matrix of the robot state. To get the robot state in the first place we need to perform state estimation using estimation techniques like Kalman filter.
In this assignment, the Extended Kalman filter is used which will be discussed in detail later in this report. One important point here to note is that estimation techniques like the Kalman filter are based on complex matrix computations which PDDL is incapable of performing. Therefore, to solve this problem, we have used a technique called semantic attachment, or external module, in planning. This technique creates a bridge between task planning and motion planning as it allows complex computations of numerical fluents by calling an external planner library.
To successfully run the project the first thing to do is to install the popf-tif planner, the details of which can be found in this repository:
If you choose to use Docker Ubuntu image then keep in mind that the planner runs on the Ubuntu-18.04 Docker Official Image which doesn’t have any graphical interface.
Therefore, to edit the planning files you need to mount a shared directory between your host system and docker image by the following command:
docker run -dit -v path/in_your/host/folder:path/in/docker_container/folder --name your_name hypothe/ai4ro2_2
However the easiest way to run the project is to use this docker image which is developed by our colleague, Marco Gabriele Fedozzi.
https://hub.docker.com/r/hypothe/ai4ro2_2
This docker image already contains all the basic setup to run the project. You just need to extract the visits_domain and visitis_module folders given along with this repository and replace them in ‘/root/ai4ro2’ directory. The visits_domain folder contains a PDDL domain, and problem files and the visits_module folder contains files to run the external planner.
Besides this, make sure that you have installed an Armadillo C++ library in your system which will be useful for complex matrix computation.
In the ‘visits_domains’ folder you will also find a ‘region_poses’ file which contains the mapping from a region to its corresponding waypoint, waypoint.txt file which contains the geometric way-point locations for the four regions to visit as well as the starting region and the landmark.txt file which contains the landmarks to localize the robot. Make sure you have changed the directory addresses for waypoint.txt and landmark.txt files inside the ‘loadSolver’ function in the ‘VisitSolver.cpp’ file. The instructions to build the external module can be found in the ‘buildInstruction.txt’ file. Once built, the planner can be run using the command:
popf3-clp -x dom.pddl prob.pddl ../visits_module/build/libVisits.so region_poses
If the external module is built correctly, the planner should run without any errors and you the planner result on the terminal like this.
We would like to extend our gratitude to:
- Dr. Antony Thomas, for providing the basic setup of the system, including domain and external planner files for us to build further on.
- Marco Gabriele Fedozzi, for helping out with the docker image loaded with all dependencies.