RoboSLAM Research

IVALab RoboSLAM Research

Our research is in SLAM for Robotics is geared towards establishing accurate, real-time realization of SLAM systems. We are interested in achieving accurate SLAM estimates during online pose estimation and map building. So far, our focus has been on the pose estimation (or localization) part of SLAM and not on the map building. Pose estimation is critical when navigating unknown environments since these same estimates are required for feedback trajectory control of the mobile robot. A poor pose estimate will only get worse under feedback, thus it is important to have the best possible estimate within a reasonable time (relative to the control feedback rate). Nevertheless, we believe that the ideas supporting rapid and robust pose estimation will translate to map building.

Key to our work has been the notion of provable approximations. SLAM can run faster if measurements are thrown out, however success and performance will vary based on which are kept and which are not. Using tools from the numerical optimization and numerical methods community, we believe it is possible to actively keep some measurements and ignore others when estimating pose during SLAM, without impacting performance. If possible to achieve, then it will be possible to enhance the runtime of slow but accurate SLAM methods in a way that provides positive benefite during controlled trajectory tracking through unknown environments. Current evidence indicates that our good features approach will indeed support such an outcome. The main tool from the numerical optimization and numerical methods community is that of submodularity, greedy optimization algorithms exploiting submodularity, and proven near-optimality outcomes associated to these algorithms. The work is a nice blend of control theory and approximation theory in support of efficient optimization algorithms.

We believe that the idea behind the work is quite general and can be applied in many parts of the SLAM pipeline where there are data-induced computational bottlenecks. In bundle-adjustment based SLAM methods, the problem size grows with time, so active mechanisms to curate the measurements are needed to preserve real-time operation. Both the front-end and back-end components suffer from this problem, thus we have embarked on a journey to better understand just how many components could be enhanced, how they could be enhanced, and how general the concept is relative to existing SLAM solutions. Our intent is to complement these solutions be leveraging their strengths and addressing their computational weaknesses, but just enough that they can function for mobile robot deployments and provide strong, high performance solutions that do not degrade in the closed-loop. A consequence of the nature of our investigation is that a lot of this work will use the good features descriptor in it, or some variant thereof, as we explore SLAM from many angles, but with the same lens.

Investigators Involved

Collaborators (Current and Former)

Former:

Former Investigators