ROS-based autonomous navigation and object recognition for a mobile manipulator operating in a warehouse environment

One of the most challenging problems nowadays is the development of robotic systems capable of carrying out increasingly problematic tasks with a high level of autonomy. In this sense, this thesis work aims to develop a software architecture for a mobile manipulator (Locobot WX250) that can: (i) recognize an obstacle along its path, and (ii) perform pick and place tasks without being aware of how the map is constructed. The mobile manipulator software architecture is implemented using ROS (Robot Operating System) to communicate within its modules. ROS represents the state of the art in middleware software and has a modular structure that can be updated, enriched, or simplified at any moment without corrupting the whole system. Furthermore, along with ROS, the manipulator’s tasks are developed using the MoveIt framework. This choice allows the use of different possible motion planning algorithms, leaving the decision on which is the most suitable depending on the end user's necessities. The obstacle awareness for the manipulator is brought through Octomap, while the algorithms picked for solving the motion planning problem are: Stochastic Trajectory Optimization for Motion Planning (STOMP), Covariant Hamiltonian Optimization for Motion Planning (CHOMP), and Open Motion Planning Library (OMPL). Thanks to the rich framework of ROS and widely available, open-source packages we were able to test various techniques for localization and mapping, path planning and obstacle avoidance. The navigation stack that is at the core of the mobile base software uses odometry and sensor data (both Lidar and RGB-D), is highly modular and can be customized for each use case. For mapping and localization we adopted RTAB-Map, a graph based SLAM algorithm, that uses real time loop-closure on previously seen locations and that makes use of all sensors available on the Locobot. Obstacle detection is achieved through Lidar and RGB-D sensors, both of which are used to construct occupancy grid costmaps that are then used by the global and local path planner to instruct the mobile base’s movements. For global planning we explore both A* and Dijkstra’s algorithms and for local planning Trajectory Rollout, Dynamic Window Approach(DWA) and Timed Elastic Band (TEB) algorithms. After a thorough comparison we selected A* for its computational speed and Timed Elastic Band for its flexibility and stability. We also explored and added a social-navigation layer for human avoidance during path planning with a Gaussian based costmap approach. Our architecture makes use of these packages to implement a search of known and unknown spaces following an ARTag request. The Locobot will reach the ARTag and perform a pick task with the manipulator. After a successful pick operation it will place the requested object in a previously chosen location. To validate our algorithm, we tested the robot both in a simulated environment, thanks to Gazebo, and in a real application using the laboratory as an adapted warehouse.