On the Development of a Generic Multi-Sensor Fusion Framework for Robust Odometry Estimation

In this work we review the design choices, the mathematical and software engineering techniques employed in the development of the ROAMFREE sensor fusion library, a general, open-source framework for pose tracking and sensor parameter self-calibration in mobile robotics. In ROAMFREE, a comprehensive logical sensor library allows to abstract from the actual sensor hardware and processing while preserving model accuracy thanks to a rich set of calibration parameters, such as biases, gains, distortion matrices and geometric placement dimensions. The modular formulation of the sensor fusion problem, which is based on state-of-the-art factor graph inference techniques, allows to handle arbitrary number of multi-rate sensors and to adapt to virtually any kind of mobile robot platform, such as Ackerman steering vehicles, quadrotor unmanned aerial vehicles, omni-directional mobile robots. Different solvers are available to target high-rate online pose tracking tasks and offline accurate trajectory smoothing and parameter calibration. The modularity, versatility and out-of-the-box functioning of the resulting framework came at the cost of an increased complexity of the software architecture, with respect to an ad-hoc implementation of a platform dependent sensor fusion algorithm, and required careful design of abstraction layers and decoupling interfaces between solvers, state variables representations and sensor error models. However, we review how a high level, clean, C++/Python API, as long as ROS interface nodes, hide the complexity of sensor fusion tasks to the end user, making ROAMFREE an ideal choice for new, and existing, mobile robot projects.