Doherty Kevin J.

No Thumbnail Available
Last Name
Doherty
First Name
Kevin J.
ORCID

Filters

2
1 1 1
2019 - 2019 1 2020 - 2023 1
2
2
Reset filters

Settings

Search Results

Now showing 1 - 2 of 2
  • Thesis
    Robust non-Gaussian semantic simultaneous localization and mapping
    (Massachusetts Institute of Technology and Woods Hole Oceanographic Institution, 2019-09) Doherty, Kevin J.
    The recent success of object detection systems motivates object-based representations for robot navigation; i.e. semantic simultaneous localization and mapping (SLAM), in which we aim to jointly estimate the pose of the robot over time as well as the location and semantic class of observed objects. A solution to the semantic SLAM problem necessarily addresses the continuous inference problems where am I? and where are the objects?, but also the discrete inference problem what are the objects?. We consider the problem of semantic SLAM under non-Gaussian uncertainty. The most prominent case in which this arises is from data association uncertainty, where we do not know with certainty what objects in the environment caused the measurement made by our sensor. The semantic class of an object can help to inform data association; a detection classified as a door is unlikely to be associated to a chair object. However, detectors are imperfect, and incorrect classification of objects can be detrimental to data association. While previous approaches seek to eliminate such measurements, we instead model the robot and landmark state uncertainty induced by data association in the hopes that new measurements may disambiguate state estimates, and that we may provide representations useful for developing decisionmaking strategies where a robot can take actions to mitigate multimodal uncertainty. The key insight we leverage is that the semantic SLAM problem with unknown data association can be reframed as a non-Gaussian inference problem. We present two solutions to the resulting problem: we first assume Gaussian measurement models, and non-Gaussianity only due to data association uncertainty. We then relax this assumption and provide a method that can cope with arbitrary non-Gaussian measurement models. We show quantitatively on both simulated and real data that both proposed methods have robustness advantages as compared to traditional solutions when data associations are uncertain.
  • Thesis
    Lifelong, learning-augmented robot navigation
    (Massachusetts Institute of Technology and Woods Hole Oceanographic Institution, 2023-02) Doherty, Kevin J. ; Leonard, John J.
    Simultaneous localization and mapping (SLAM) is the process by which a robot constructs a global model of an environment from local observations of it; this is a fundamental perceptual capability supporting planning, navigation, and control. We are interested in improving the expressiveness and operational longevity of SLAM systems. In particular, we are interested in leveraging state-of-the-art machine learning methods for object detection to augment the maps robots can build with object-level semantic information. To do so, a robot must combine continuous geometric information about its trajectory and object locations with discrete semantic information about object classes. This problem is complicated by the fact that object detection techniques are often unreliable in novel environments, introducing outliers and making it difficult to determine the correspondence between detected objects and mapped landmarks. For robust long-term navigation, a robot must contend with these discrete sources of ambiguity. Finally, even when measurements are not corrupted by outliers, long-term SLAM remains a challenging computational problem: typical solution methods rely on local optimization techniques that require a good “initial guess,” and whose computational expense grows as measurements accumulate. The first contribution of this thesis addresses the problem of inference for hybrid probabilistic models, i.e., models containing both discrete and continuous states we would like to estimate. These problems frequently arise when modeling e.g., outlier contamination (where binary variables indicate whether a measurement is corrupted), or when performing object-level mapping (where discrete variables may represent measurement-landmark correspondence or object categories). The former application is crucial for designing more robust perception systems. The latter application is especially important for enabling robots to construct semantic maps; that is, maps containing objects whose states are a mixture of continuous (geometric) information and (discrete) categorical information (such as class labels). The second contribution of this thesis is, a novel spectral initialization method which is efficient to compute, easy to implement, and admits the first formal performance guarantees for a SLAM initialization method. The final contribution of this thesis aims to curtail the growing computational expense of long-term SLAM. In particular, we propose an efficient algorithm for graph sparsification capable of reducing the computational burden of SLAM methods without significantly degrading SLAM solution quality. Taken together, these contributions improve the robustness and efficiency of robot perception approaches in the lifelong setting.