Robotics and Intelligent Machines Center, College of Comp

uting, Georgia Institute of Technology

Modern mobile robotic systems are required to exhibit high levels of autonomy, and in particular to be able to localize and navigate with high-precision in unknown environments. Mapping these environments at the same time, using onboard sensors such as monocular and stereo cameras, leads to the well known simultaneous localization and mapping (SLAM) problem which is commonly solved using bundle adjustment techniques. While considerable progress has been made in recent years to efficiently incorporate incoming observations from onboard sensors, high-rate performance is still a challenge in many cases. This is in particular the case when operating in large-scale environments over long time periods and in presence of loop closures, i.e. re-observing 3D points at different time instances, potentially from different viewpoints.

In this talk, I will present approaches that address these challenges, mainly focusing on an incremental and computationally efficient method for bundle adjustment that substantially reduces the involved computational cost, compared to state-of-the-art bundle adjustment techniques. The method - incremental light bundle adjustment (iLBA) - incorporates the following two key components. First, the observed 3D points (structure) are algebraically eliminated and the cost function is formulated using multi-view constraints instead of the projection equations, leading to a reduced number of variables in the optimization. These include only the pose (or navigation state) variables and do not include structure parameters. If required, the observed 3D points, or any part of them, can be reconstructed based on the optimized poses of the robot. The second component is the recently developed incremental smoothing approach, which adaptively identifies the variables that need to be recomputed at each step. The optimization problem is formulated in terms of a graphical model, a factor graph, and it is shown how to perform an efficient inference that typically involves only a fraction of the pose variables and fully exploits sparsity of the system.

In the final part of the talk, I will discuss an extension of iLBA to multi-robot localization and outline an approach that parallelizes computations to maintain high-rate performance also in the presence of loop closure observations. Since this approach uses the same inference engine as iLBA, the two methods can be combined into a single framework.