Mobile manipulation using the Boston Dynamics Spot robot
Basic Information
Students: Marko Koskinen, Jaakko Tuomisto, Niklas Lindgren, Chen Wang
Project manager: Marko Koskinen
Instructor: Nikita Kostin
Starting date: 11.1.2022
Completion date: 3.6.2022
Abstract
The purpose of our project was to develop software for the Boston Dynamics Spot robot and test it in a practical environment. The goal was to implement algorithms that would enable the robot to traverse through doors independently of operators. Even though Boston Dynamics provided algorithms for opening doors automatically, they were not designed for fully autonomous operation. Additionally, they had a very low success rate when opening the university doors.
To autonomously open doors, we needed to develop solutions for navigation and mapping, object detection, and the actual door opening movements. For the navigation, we built a custom path finding implementation. For mapping, we successfully used an open source package called RTAB-Map. Object detection was done with a deep learning model called YOLOv5. To train the object detection model, we collected data from the university campus featuring different handles and doors. The model was trained using Aalto University’s Triton computing cluster. The actual door opening movements were planned by us, and executed by the robot via the Boston Dynamics SDK (software development kit).
By the end the of the project, we had implemented a working solution for mapping, navigation, door detection, and handle detection. In addition, we implemented additional features in the Clearpath Robotics ROS Spot driver for operating the arm and decreasing the network bandwidth used by sensor information. These components together formed a working solution for opening doors without operator input. The solution was tested with different doors around the university campus and it had a high success rate.
