Author: Anna V. Petrovskaya
Publisher: Stanford University
Published: 2011
Total Pages: 226
ISBN-13:
DOWNLOAD EBOOK →
Reliable perception is required in order for robots to operate safely in unpredictable and complex human environments. However, reliability of perceptual inference algorithms has been poorly studied so far. These algorithms capture uncertain knowledge about the world in the form of probabilistic belief distributions. A number of Monte Carlo and deterministic approaches have been developed, but their efficiency depends on the degree of smoothness of the beliefs. In the real world, the smoothness assumption often fails, leading to unreliable perceptual inference results. Motivated by concrete robotics problems, we propose two novel perceptual inference algorithms that explicitly consider local non-smoothness of beliefs and adapt to it. Both of these algorithms fall into the category of iterative divide-and-conquer methods and hence scale logarithmically with desired accuracy. The first algorithm is termed Scaling Series. It is an iterative Monte Carlo technique coupled with annealing. Local non-smoothness is accounted for by sampling strategy and by annealing schedule. The second algorithm is termed GRAB, which stands for Guaranteed Recursive Adaptive Bounding. GRAB is an iterative adaptive grid algorithm, which relies on bounds. In this case, local non-smoothness is captured in terms of local bounds and grid resolution. Scaling Series works well for beliefs with sharp transitions, but without many discontinuities. GRAB is most appropriate for beliefs with many discontinuities. Both of these algorithms far outperform the prior art in terms of reliability, efficiency, and accuracy. GRAB is also able to guarantee that a quality approximation of the belief is produced. The proposed algorithms are evaluated on a diverse set of real robotics problems: tactile perception, autonomous driving, and mobile manipulation. In tactile perception, we localize objects in 3D starting with very high initial uncertainty and estimating all 6 degrees of freedom. The localization is performed based on tactile sensory data. Using Scaling Series, we obtain highly accurate and reliable results in under 1 second. Improved tactile object localization contributes to manufacturing applications, where tactile perception is widely used for workpiece localization. It also enables robotic applications in situations where vision can be obstructed, such as rescue robotics and underwater robotics. In autonomous driving, we detect and track vehicles in the vicinity of the robot based on 2D and 3D laser range finders. In addition to estimating position and velocity of vehicles, we also model and estimate their geometric shape. The geometric model leads to highly accurate estimates of pose and velocity for each vehicle. It also greatly simplifies association of data, which are often split up into separate clusters due to occlusion. The proposed Scaling Series algorithm greatly improves reliability and ensures that the problem is solved within tight real time constraints of autonomous driving. In mobile manipulation, we achieve highly accurate robot localization based on commonly used 2D laser range finders using the GRAB algorithm. We show that the high accuracy allows robots to navigate in tight spaces and manipulate objects without having to sense them directly. We demonstrate our approach on the example of simultaneous building navigation, door handle manipulation, and door opening. We also propose hybrid environment models, which combine high resolution polygons for objects of interest with low resolution occupancy grid representations for the rest of the environment. High accuracy indoor localization contributes directly to home/office mobile robotics as well as to future robotics applications in construction, inspection, and maintenance of buildings. Based on the success of the proposed perceptual inference algorithms in the concrete robotics problems, it is our hope that this thesis will serve as a starting point for further development of highly reliable perceptual inference methods.