There may arise two kinds of challenges in the problem of mobile robot localization; (i) a robot may have an a priori map of its environment, in which case the localization problem boils down to estimating the robot pose relative to a global frame or (ii) no a priori map information is given, in which case a robot may have to estimate a model of its environment and localize within it. In the case of a known map, simultaneous planning while localizing is a crucial ability for operating under uncertainty. We first address this problem by designing a method to dynamically replan while the localization uncertainty or environment map is updated. Extensive simulations are conducted to compare the proposed method with the performance of FIRM (Feedback-based Information RoadMap). However, a shortcoming of this method is its reliance on a Gaussian assumption for the Probability Density Function (pdf) on the robot state. This assumption may be violated during autonomous operation when a robot visits parts of the environment which appear similar to others. Such situations lead to ambiguity in data association between what is seen and the robot's map leading to a non-Gaussian pdf on the robot state. We address this challenge by developing a motion planning method to resolve situations where ambiguous data associations result in a multimodal hypothesis on the robot state. A Receding Horizon approach is developed, to plan actions that sequentially disambiguate a multimodal belief to achieve tight localization on the correct pose in finite time. In our method, disambiguation is achieved through active data associations by picking target states in the map which allow distinctive information to be observed for each belief mode and creating local feedback controllers to visit the targets. Experiments are conducted for a kidnapped physical ground robot operating in an artificial maze-like environment.
The hardest challenge arises when no a priori information is present. In longterm tasks where a robot must drive for long durations before closing loops, our goal is to minimize the localization error growth rate such that; (i) accurate data associations can be made for loop closure, or (ii) in cases where loop closure is not possible, the localization error stays limited within some desired bounds. We analyze this problem and show that accurate heading estimation is key to limiting localization error drift. We make three contributions in this domain. First we present a method for accurate long-term localization using absolute orientation measurements and analyze the underlying structure of the SLAM problem and how it is affected by unbiased heading measurements. We show that consistent estimates over a 100km trajectory are possible and that the error growth rate can be controlled with active data acquisition. Then we study the more general problem when orientation measurements may not be present and develop a SLAM technique to separate orientation and position estimation. We show that our method's accuracy degrades gracefully compared to the standard non-linear optimization based SLAM approach and avoids catastrophic failures which may occur due a bad initial guess in non-linear optimization. Finally we take our understanding of orientation sensing into the physical world and demonstrate a 2D SLAM technique that leverages absolute orientation sensing based on naturally occurring structural cues. We demonstrate our method using both high-fidelity simulations and a real-world experiment in a 66, 000 square foot warehouse. Empirical studies show that maps generated by our approach never suffer catastrophic failure, whereas existing scan matching based SLAM methods fail ? 50% of the time.
- Chakravorty, Suman Associate Professor