Longterm Discovery and Modeling of Temporal Phenomena to Support Robotic Service Behaviors

Due to the dynamic nature of real-world environments, an assistive robot must be able to adapt its services to how the environment is changing over time if it is to provide effective, continued support over the course of long-term deployments. As a course of research, I have investigated how a robotic agent can actively learn long-term models of its environment while jointly leveraging such models to perform more efficient service tasks. Specifically, the stated goal has been to improve the supportive capabilities of an assistive robot by integrating continual learning of temporal phenomena with exploration and routine task behavior. This research has resulted in two main contributions: a framework facilitating the semi-supervised learning of object classifiers as well as an approach for jointly planning observation gathering and task scheduling, specifically in the context of time-dependent delivery problems. For my first contribution, I developed a data gathering framework by which human operators can direct the mostly autonomous learning of object classifiers with little manual effort. Such work was aimed to support various assistive robotic tasks depending on visual detection of specialized objects not well represented in large publicly available data sets. While visual detectors have seen significant gains in performance, such gains do not translate when sufficient training data is unavailable. To address this, I developed an interface using the RVIZ visualizer by which a human operator can direct a robot to autonomously collect a visual data set of new objects through image capture from various sampled views. This is performed with only a couple quick interactions with an easy to use interface on the part of the human operator. After a quick setup, the robot will sample various spatial viewpoints that it navigates to for image capture. Leveraging related work in the form of a transfer learning technique known as fine-tuning, a visual classifier first trained on out of domain images is then quickly retrained for the relevant objects with only having to capture a selection of images.Additionally, only collected images found to result in increasing classifier performance on validation sets are trained on, resulting in small dataset storage requirements for future retraining. This approach was shown to significantly improve classification performance on in domain objects when compared to using available pre-constructed datasets. Additionally, this iterative view sampling method was shown to find a good balance between classifier performance and data storage constraints. This work resulted in the paper “Learning Object Classifiers with Limited Human Supervision on a Physical Robot”, which I presented at the 2018 IEEE International Conference on Robotic Computing (IRC).For my second major contribution, I developed a planning approach for Pickup and Delivery Problems with customer availabilities that are stochastic, time-dependent, and dynamically revealed. As traditional policy-based methods do not scale to problems of the size addressed in this work, I developed a heuristic-based dynamic programming planner that draws inspiration from related work applied to Orienteering Problems with time-dependent rewards. Specifically, this planner uses dynamic programming to find a maximum profit path linking observation, task servicing (delivery), and movement actions in a spatio-temporal graph decomposition of the environment. Observation actions are heuristically motivated according to the projected decrease in entropy (uncertainty) of the temporal model when a given observation is integrated, and delivery actions are heuristically motivated according with current model uncertainty based on most recent observations. This planning approach was integrated with a temporal modeling approach combining related work in long-term temporal modeling and short-term temporal persistence modeling using Bayesian filtering to provide estimates of delivery task success. In this way, observational actions to support more accurate temporal models of the environment are integrated alongside delivery actions in such a way as to maximize task efficiency by lowering the chance of delivery failures due to environmental dynamics. This combined delivery and observation planning approach was shown to have vast computational benefits over policy-based approaches, outperforming them on simulated office delivery scenarios when given similar computation. This method for including observations similarly greatly outperformed baseline approaches that simplistically port dynamic programming approaches to Orienteering Problem-based planning.