Counter-Hypothetical Evidential Reasoning for Mobile Manipulation Robots

Abstract

Although robots can perform in structured environments, they struggle to perceive and operate within cluttered, dynamic, and previously unseen settings. Nonparametric Bayesian inference has the potential to address these problems caused by uncertainty and reason over nonlinear high dimensional states. However, sampling-based Bayesian inference methods are susceptible to mode collapse of the belief distribution, where the inference process incorrectly converges to a single region of the state space. Inference with more samples can improve the ability to represent the state space fully, but it is often not feasible due to computational and resource constraints in robotics domains.

This dissertation introduces a counter-hypothetical approach to evidential reasoning for addressing the problems of mode collapse in nonparametric Bayesian inference. Evidential reasoning, in the context of nonparametric Bayesian inference, allows us to explicitly model likelihood, ambiguity, and doubt in the underlying belief of the distribution. With these more delineated measurements of belief, we no longer need to infer quantities of ambiguity and doubt from likelihood weightings alone. We demonstrate this extension can enable nonparametric Bayesian inference to sample over high-dimensional state spaces with more robustness to particle deprivation.

We begin by introducing the Counter-Hypothetical Particle Filter, CH-PF, to overcome mode collapse when tracking rigid objects from monocular video observations. Previous methods predict failures during inference based on the likelihood function. We observe that low likelihood weightings for a given hypothesis can be attributed to error in the pose or ambiguity in the observation. For this reason, we present the counter-hypothetical likelihood function to estimate doubt independently of likelihood or ambiguity. The counter-hypothetical particle filter quantifies the evidence that supports a hypothesis and refutes the hypothesis. This independent and explicit modeling of doubt through a proposed counter-hypothetical likelihood function enables the filter to better detect failure modes and adaptively redistribute probability mass to a null hypothesis.

To better evaluate the performance of our methods on tracking high-dimensional states under heavy occlusion, we present a benchmark dataset. The Progress LUMBER (Looking Upon a Moving BipEdal Robot) Dataset contains 100 sequences of a bipedal humanoid robot Digit moving within highly obstructed scenes. The annotations require no external markers to label the pose of 29 links. The occlusions featured in the dataset make it unique in its representation of real-world environments that humanoid mobile manipulators may face in practical scenarios.

Extending counter-hypothetical reasoning to higher dimensional systems, we present Weighted And Graphical Evidential Reasoning for Differentiable Nonparametric Belief Propagation, (WAGER-DNBP). This method models evidential reasoning within a differentiable nonparametric belief propagation algorithm. WAGER-DNBP not only learns the unary and pairwise potentials via labeled tracking data but also the counter-hypothetical likelihood. We then use inconsistencies between a given hypothesis's likelihood and counter-hypothetical likelihood scores and observation to estimate ambiguity. WAGER-DNBP then uses these measurements of ambiguity to determine which observations within the factor graph should carry more weight in the posterior belief distribution. We validate our method on the Progress LUMBER Dataset to show that the explicit modeling of ambiguity and doubt within WAGER-DNBP can enable it to recover from particle deprivation more efficiently.