Efficient Belief Propagation for Perception and Manipulation in Clutter

Abstract

Autonomous service robots are required to perform tasks in common human indoor environments. To achieve goals associated with these tasks, the robot should continually perceive, reason its environment, and plan to manipulate objects, which we term as goal-directed manipulation. Perception remains the most challenging aspect of all stages, as common indoor environments typically pose problems in recognizing objects under inherent occlusions with physical interactions among themselves. Despite recent progress in the field of robot perception, accommodating perceptual uncertainty due to partial observations remains challenging and needs to be addressed to achieve the desired autonomy.

In this dissertation, we address the problem of perception under uncertainty for robot manipulation in cluttered environments using generative inference methods. Specifically, we aim to enable robots to perceive partially observable environments by maintaining an approximate probability distribution as a belief over possible scene hypotheses. This belief representation captures uncertainty resulting from inter-object occlusions and physical interactions, which are inherently present in clutterred indoor environments. The research efforts presented in this thesis are towards developing appropriate state representations and inference techniques to generate and maintain such belief over contextually plausible scene states. We focus on providing the following features to generative inference while addressing the challenges due to occlusions: 1) generating and maintaining plausible scene hypotheses, 2) reducing the inference search space that typically grows exponentially with respect to the number of objects in a scene, 3) preserving scene hypotheses over continual observations.

To generate and maintain plausible scene hypotheses, we propose physics informed scene estimation methods that combine a Newtonian physics engine within a particle based generative inference framework. The proposed variants of our method with and without a Monte Carlo step showed promising results on generating and maintaining plausible hypotheses under complete occlusions. We show that estimating such scenarios would not be possible by the commonly adopted 3D registration methods without the notion of a physical context that our method provides.

To scale up the context informed inference to accommodate a larger number of objects, we describe a factorization of scene state into object and object-parts to perform collaborative particle-based inference. This resulted in the Pull Message Passing for Nonparametric Belief Propagation (PMPNBP) algorithm that caters to the demands of the high-dimensional multimodal nature of cluttered scenes while being computationally tractable. We demonstrate that PMPNBP is orders of magnitude faster than the state-of-the-art Nonparametric Belief Propagation method. Additionally, we show that PMPNBP successfully estimates poses of articulated objects under various simulated occlusion scenarios.

To extend our PMPNBP algorithm for tracking object states over continuous observations, we explore ways to propose and preserve hypotheses effectively over time. This resulted in an augmentation-selection method, where hypotheses are drawn from various proposals followed by the selection of a subset using PMPNBP that explained the current state of the objects. We discuss and analyze our augmentation-selection method with its counterparts in belief propagation literature. Furthermore, we develop an inference pipeline for pose estimation and tracking of articulated objects in clutter. In this pipeline, the message passing module with the augmentation-selection method is informed by segmentation heatmaps from a trained neural network. In our experiments, we show that our proposed pipeline can effectively maintain belief and track articulated objects over a sequence of observations under occlusion.