The current landscape of modern field robotics confronts a notable expansionchallenge, primarily stemming from the proliferation of diverse platforms, vast data vol- umes, sophisticated algorithms, and powerful computational resources. However, the progress of robotics beyond its conventional role as a mere observational tool is im- peded by limited autonomy and undeveloped data inference capabilities. The scarcity of labeled data poses a significant hurdle in the application of machine learning to novel tasks in the space of image comprehension, which incur exorbitant costs for manual labeling. The traditional approach of relying on human annotators at workstations for prolonged periods of time exhibits several limitations, including subjective consis- tency and human factors concerns. Conversely, the adoption of human-centric labeling tools presents a viable solution to address these challenges by deploying human experts in the field to perform data labeling, thereby enabling swift inspection of the sensed environment while eliminating data representation as a bottleneck. This approach fa- cilitates the rapid annotation of data in a sensor-agnostic manner, culminating in the generation of large-scale, high-quality datasets within feasible timeframes by leveraging comparatively small amounts of expert-derived knowledge to improve data labels over all.

The adoption of Human-Robot (HR) teaming continues to increase within many high-risk fields (e.g., space exploration, medical treatment). This work solves an important problem in HR teaming by learning to interpret Mental Workload (MW) from human passive biosignals in the context of human performance. Using a previously designed experiment, we analyze GP kernels for human performance estimation. GP models help limit designer bias and provide prediction confidence intervals. This analysis offers the following contributions to this field: a heuristic to help understand which biosignals are informative, an evaluation of data, and a comparison of GP kernels. The experiment showed that smooth kernels yield lower performance prediction Root Mean Squared Error (RMSE) and Standard Error of the Mean (SEM) for each performance metric considered. As a result of this work, performance prediction model designers will have a guide for improving HR systems, passive MW monitoring, and performance estimation of HR teaming.

In order to harness the potential of deep learning and computer object recogni-tion in practical environments, a substantial collection of interesting features is essential. However, the creation of labeled image datasets is a significant challenge, hindering the broader adoption and autonomy by farmers, land managers, and ecologists. To address this challenge, we present a data collection, annotation, and pro- cessing pipeline utilizing Unmanned Aerial Vehicle (UAV) based optical sensing. Bird- sEye empowers non-deep learning experts to train, maintain, and deploy sophisticated computer vision methods on their own local land environments by substituting the need for aerial image feature identification with terrain based observation by subject matter experts. These annotated ground observations are then used to identify relevant image sections within UAV captured imagery. By facilitating the rapid generation of labeled datasets, our approach can iden- tify and characterize a diverse array of land and plant conditions. This has significant applications in areas such as disease monitoring, vegetation and pest identification, and precision treatments.

Precise ecological assessment of forest environments relies on acquiring dense, accurate spatial reconstructions. Point clouds serve as the leading form of three dimensional (3D) forest representation, and have been successfully used to recover important structural parameters of trees, namely diameter at breast height (DBH). Point cloud generation has traditionally relied on photogrammetric or Light Detection and Ranging (LiDAR) sensing modalities. Neural Radiance Fields (NeRFs), a recent innovation in artificial intelligence and computer vision, enable point cloud generation from a neural network trained on a sparse set of 2D images. This thesis presents an evaluation of three point cloud generation methods for the purpose of tree localization and DBH estimation in a coastal redwood forest. The first method uses lidar-inertial Simultaneous Localization and Mapping (SLAM) and serves as the state-of-the-art comparison. The two remaining methods generate neural-implicit scene geometry by training a NeRF on visual and SLAM data sourced from mobile phone and robot platforms. The results present an exciting avenue for rapid ecological assessment of forests environments using community-sourced mobile phone imagery as an alternative to expensive lidar sensing.