Distributed Inference for Robotic Perception and Planning Under Uncertainty

Abstract

Autonomous robots promise to improve human productivity and quality of life by assisting in our homes, factories, and labs to automate the dull, dirty, and dangerous. To achieve these long-promised versatile robot assistants, robots must be able to operate robustly in unstructured real-world environments. Environments made for humans remain challenging for autonomous robots due to their highly unstructured nature which arises from environmental occlusions, dynamic environments, and the diversity of possible objects a robot might encounter. These challenges result in a plurality of competing hypotheses present at all levels of the system. Modelling the uncertainty inherent to the robotic system is a crucial capability to enable robust operation in unstructured environments.

This dissertation considers the problem of scalable, robust operation under uncertainty through distributed probabilistic inference. To handle the intractable nature of these problems, distributed probabilistic inference decomposes high-dimensional problems into simpler, parallelizable subproblems. These are represented as probabilistic graphical models and solved via message passing. Further, the resulting distributions must encompass arbitrary, multi-modal uncertainty which results from competing hypotheses and noisy estimates. We employ nonparametric distributions for their flexible representational capabilities. We present novel approaches leveraging these insights and demonstrate their application to robotic perception and planning problems.

First, we consider the problem of articulated object localization in cluttered scenes towards robot manipulation of hand-tools. We take a parts-based approach, modelling object geometries by decomposing them into their component parts. We employ Nonparametric Belief Propagation to perform distributed inference over the resulting graphical model. A learned observation likelihood is leveraged alongside object geometry in order to infer the belief over the part poses. We demonstrate that the proposed method is robust to challenging observations with heavy occlusion on a custom dataset.

Second, we turn to the task of robotic planning for object manipulation. Many tasks in robotic manipulation are described by goals that are intractable to model explicitly (e.g. stable grasping or user preferences). We present a planning framework which considers uncertainty in the goal specification. To accomplish this, we consider robotic planning through the lens of probabilistic inference, modelling both the trajectory and goal as distributions. We propose a novel differentiable loss over arbitrary nonparametric goals which is demonstrated on high-dimensional robotic manipulation tasks of grasping and placement.

Third, we extend the problem of planning as inference to the high-dimensional problem of multi-robot coordination. We propose Stein Variational Belief Propagation, an algorithm for performing inference over graphical models. We show that the proposed algorithm is more effective at representing the underlying distribution than sampling-based baselines. We demonstrate the capability of this method to solve challenging, dynamic problems in robotics through multi-robot coordination experiments.

The promise of robotics coupled with the open challenges that remain in the field as described in this dissertation highlight the immediate need to train the next generation of diverse talent with expertise in robotics. Towards this objective, this dissertation formalizes recent trends in robotics undergraduate education. We present a modular introductory robotics curriculum which involves programming a custom robotic platform appropriate for undergraduate instructional use. Finally, we suggest best practices for teaching robotics as a discipline at the undergraduate level based on lessons learned teaching the described course at the University of Michigan and three partner institutions to over 100 students over the past three years.