David Wettergreen

Advances in machine learning (ML) over the past decade have resulted in a proliferation of algorithmic applications for encoding, characterizing, and acting on complex data that may contain numerous multidimensional features. Recently, the emergence of deep-learning models trained across large datasets has created a new paradigm for ML in the form of Foundation Models (FMs). FMs are programs trained on large and broad datasets with an extensive number of parameters. Once built, these extremely powerful, flexible models can be utilized in less resource-intensive ways to build a variety of different downstream applications that can integrate previously disparate, multimodal data. The development of these applications can be done rapidly and with a much lower demand for ML expertise. Additionally, the necessary infrastructure and models themselves are already established within agencies such as NASA and ESA. At NASA, this work extends across several divisions of the Science Mission Directorate. Examples include the NASA Goddard and INDUS Large Language Models and the Prithvi Geospatial Foundation Model. Furthermore, ESA initiatives to bring FMs to Earth observations have led to the development of TerraMind. In February 2025, a workshop was held by NASA Ames Research Center and the SETI Institute to explore the potential of FMs in astrobiological research and identify the steps necessary to build and utilize such a model or models. Here, we share the findings and recommendations of that workshop and describe clear near-term and future opportunities in the development of a FM (or Models) for astrobiology applications. These applications would include a biosignature or life characterization task, a mission development and operations task, and a natural language task for integrating and supporting astrobiology research needs.

In disaster response scenarios, heterogeneous robot teams can improve information gathering by leveraging diverse sensing and mobility capabilities. However, many multi-agent exploration approaches assume predetermined teams or rely on expert intuition for team selection, which may overlook better team compositions for specific information gathering tasks. This paper introduces a novel method for robot team selection based on spatial and spectral characteristics of the search domain. We define a "skill measure" for each robot and a "spectral demand" for the search region, integrating these with spatial information to guide the selection of a robot team tailored to the task. Our approach systematically matches robot capabilities to tasks, improving the effectiveness of information collection efforts. Through experiments on synthetic and real-world disaster site data, we show that our method outperforms baseline selection strategies and aligns closely with expert-selected teams in terms of coverage.

Advances in machine learning over the past decade have resulted in a proliferation of algorithmic applications for encoding, characterizing, and acting on complex data that may contain many high dimensional features. Recently, the emergence of deep-learning models trained across very large datasets has created a new paradigm for machine learning in the form of Foundation Models. Foundation Models are programs trained on very large and broad datasets with an extensive number of parameters. Once built, these powerful, and flexible, models can be utilized in less resource-intensive ways to build many different, downstream applications that can integrate previously disparate, multimodal data. The development of these applications can be done rapidly and with a much lower demand for machine learning expertise. And the necessary infrastructure and models themselves are already being established within agencies such as NASA and ESA. At NASA this work is across several divisions of the Science Mission Directorate including the NASA Goddard and INDUS Large Language Models and the Prithvi Geospatial Foundation Model. And ESA initiatives to bring Foundation Models to Earth observations has led to the development of TerraMind. A workshop was held by the NASA Ames Research Center and the SETI Institute, in February 2025, to investigate the potential of Foundation Models for astrobiological research and to determine what steps would be needed to build and utilize such a model or models. This paper shares the findings and recommendations of that workshop, and describes clear near-term, and future opportunities in the development of a Foundation Model (or Models) for astrobiology applications. These applications would include a biosignature, or life characterization, task, a mission development and operations task, and a natural language task for integrating and supporting astrobiology research needs.

Earth observing satellites are powerful tools for collecting scientific information about our planet, however they have limitations: they cannot easily deviate from their orbital trajectories, their sensors have a limited field of view, and pointing and operating these sensors can take a large amount of the spacecraft's resources. It is important for these satellites to optimize the data they collect and include only the most important or informative measurements. Dynamic targeting is an emerging concept in which satellite resources and data from a lookahead instrument are used to intelligently reconfigure and point a primary instrument. Simulation studies have shown that dynamic targeting increases the amount of scientific information gathered versus conventional sampling strategies. In this work, we present two different learning-based approaches to dynamic targeting, using reinforcement and imitation learning, respectively. These learning methods build on a dynamic programming solution to plan a sequence of sampling locations. We evaluate our approaches against existing heuristic methods for dynamic targeting, showing the benefits of using learning for this application. Imitation learning performs on average 10.0\% better than the best heuristic method, while reinforcement learning performs on average 13.7\% better. We also show that both learning methods can be trained effectively with small amounts of data.

We develop a coverage approach for heterogeneous agents that leverages the different sensing and motion capabilities of a team. Coverage performance is measured using ergodicity, which when optimized balances exploitation versus exploration, where areas of interest are indicated with an information metric. Prior work uses spectral decomposition of a spatial map of information to guide a set of heterogeneous agents, each with different sensor and motion models, to optimize coverage. This work leverages wavelet transforms to decompose the information map rather than the Fourier transform typically applied to ergodic search and demonstrates the importance of selecting a suitable wavelet family to use, based on the information map being explored. Further a sequence of wavelets is used for decomposition to overcome dependency on selecting one suitable wavelet family. Our experimental results show that using wavelet families well-suited to the specific information map for information map decomposition leads to, on average, 43% improvement over a baseline method in terms of a standard coverage metric (ergodicity), while using a wellsequenced set of wavelets for decomposition leads to a 65% improvement in coverage performance across multiple types of information maps.

Robots are often employed in hazardous or inaccessible environments, such as disaster sites, extraterrestrial terrains, agricultural fields, and ocean floors. Autonomous operation is crucial in these scenarios to reduce reliance on human operators and enable real-time decision-making. However, robots must balance multiple, often conflicting, objectives. These objectives are subject to change based on new data or evolving conditions. This paper presents a novel approach to dynamic multi-objective trajectory planning. The proposed method leverages the boundary intersection decomposition technique to adaptively plan trajectories that balance multiple evolving objectives. Our approach ensures efficient and effective exploration by continuously optimizing the trade-offs between changing objectives. We show that our method performs on average 34 % better in terms of solution quality on the dynamic multi-objective trajectory planning problem as compared to prior work.

Abstract A strategy for planetary exploration using a rover capable of science autonomy is presented. We encoded into a rover a set of driving hypotheses pertaining to the geologic origin of a field site and equipped the rover with the instrumentation needed to measure the observables related to the hypotheses, as well as the software tools to analyze them to a relatively high level of confidence. We investigated the effects of different exploration strategies that make use of rover science autonomy and compared the operational efficiency and science yield of three geological exploration scenarios: (1) standard human-directed exploration, (2) rover-directed exploration, and (3) astronaut/rover collaborative exploration. We show that exploration with a rover capable of science autonomy is operationally more efficient than the human-directed strategy, resulting in higher rates of data collection and hence a greater science yield per command cycle. Additionally, we explored and developed astronaut/rover collaborative exploration strategies and present a basic framework for effective planetary exploration that leverages the expertise of a science team, the efficiency of a science-autonomous rover, and the contextual abilities of astronauts.

Abstract A system for rapid analysis of spectroscopy data with emphasis on planetary surfaces, both imaging and single-spectrum data, is described. The system, called Tetracorder, is commanded by an expert system developed by expert spectroscopists. The Tetracorder and the expert system apply multiple algorithms to analyze a spectrum in segments, leveraging the advantages of each spectral region’s sensitivity to detecting different compounds, whether solid, liquid, or gas. The algorithms compare measured spectra to the spectral properties of materials in spectral libraries. The libraries include pure minerals, mineral mixtures that include areal mixtures, intimate mixtures, coatings, and molecular mixtures and other compounds such as organics, vegetation, liquids, and gases. Absorption bands of a particulate surface change shape with grain size, and shape changes are used in some cases to constrain grain size of each component in the surface. The different algorithm results are compared for each spectral region, and specific material composition and average grain size (when possible) are identified. The system is operational analyzing real-time data on a new generation of rovers for future planetary missions, as well as identifying materials using an imaging spectrometer on the International Space Station. Four abundance models are presented, each with increasing sophistication, that are computationally fast on imaging spectrometer data and use Tetracorder identifications to produce maps of mineral abundances. A fifth full radiative model that includes multilayer surfaces is presented but is computationally intensive. The system is open source and available on GitHub.

In disaster response scenarios, deploying robotic teams effectively is crucial for improving situational awareness and enhancing search and rescue operations. The use of robots in search and rescue has been studied but the question of where to start robot deployments has not been addressed. This work addresses the problem of optimally selecting starting locations for robots with heterogeneous capabilities-those equipped with different sensing and motion modalities-by formulating a joint optimization problem. To determine start locations, this work adds a constraint to the ergodic optimization framework whose minimum assigns robots to start locations. This becomes a little more challenging when the robots are heterogeneous - equipped with different sensing and motion modalities - because not all robots start at the same location, and a more complex adaptation of the aforementioned constraint is applied. Our method assumes access to potential starting locations, which can be obtained from expert knowledge or aerial imagery. We experimentally evaluate the efficacy of our joint optimization approach by comparing it to baseline methods that use fixed starting locations for all robots. Our experimental results show significant gains in coverage performance, with average improvements of 35.98% on synthetic data and 31.91 % on real-world data for homogeneous and heterogeneous teams, in terms of the ergodic metric.

Coral reefs are an important ecosystem to the local communities and indigenous wildlife that rely on them. However, reefs have greatly degraded in recent decades with the remaining at increasing risk of loss. Quantitatively mapping these reefs would provide a resource for us to monitor changes and understand their health. We explore methods leveraging limited spectral data and resources for efficient global scale modeling of coral reefs. We then evaluate performance on a Deep Neural Network and our previously developed Deep Conditional Dirichlet Model. Regions of high uncertainty based on the model output prediction are used to determine informative in situ sampling. An ergodic planner is implemented to generate a path through these regions to acquire samples that best improve the coral map. The result is a resource efficient learning based pipeline that augments existing spectral data and maps coral reefs globally to improve our understanding of their condition.

Aerial imagery is increasingly used in Earth science and natural resource management as a complement to labor-intensive ground-based surveys. Aerial systems can collect overlapping images that provide multiple views of each location from different perspectives. However, most prediction approaches (e.g. for tree species classification) use a single, synthesized top-down "orthomosaic" image as input that contains little to no information about the vertical aspects of objects and may include processing artifacts. We propose an alternate approach that generates predictions directly on the raw images and accurately maps these predictions into geospatial coordinates using semantic meshes. This method$\unicode{x2013}$released as a user-friendly open-source toolkit$\unicode{x2013}$enables analysts to use the highest quality data for predictions, capture information about the sides of objects, and leverage multiple viewpoints of each location for added robustness. We demonstrate the value of this approach on a new benchmark dataset of four forest sites in the western U.S. that consists of drone images, photogrammetry results, predicted tree locations, and species classification data derived from manual surveys. We show that our proposed multiview method improves classification accuracy from 53% to 75% relative to an orthomosaic baseline on a challenging cross-site tree species classification task.

The geometric accuracy of digital elevation models built from neural radiance fields (NeRFs) is assessed using stereo pairs collected during a simulated rover traverse under lunar polar lighting conditions by comparison to multi-view stereo reconstruction. While NeRF-based methods are more sensitive to the viewpoints in the training data and produce more artifacts on the edges of the scene, they are capable of producing denser models in occluded regions with limited additional error when the light source is not visible in the cameras. With a visible light source, the NeRF models are incapable of correctly learning scene geometry, though rendered images still appear to be decent. This trend is mitigated somewhat by using depth supervision, though this method elsewhere produces higher amounts of error. Since the volumetric rendering used by NeRF relies on probabilistic reasoning along the ray used to observe the scene, the standard deviation and gradient of the cumulative distribution function can be used as indicators of how sharply a NeRF model resolves a surface and are correlated with height error.

The MoonRanger micro-rover will explore the South Pole of the Moon in search of water ice. Its small size, and therefore limited power and communication capabilities, necessitates autonomous navigation rather than teleoperation. MoonRanger’s size and limited computing also introduce unique challenges to designing robust methods of navigation. This work describes the trajectory planning algorithms and architecture that will allow MoonRanger to operate safely and efficiently on the surface of the Moon. We verify that these methods meet mission requirements through tests in simulation and with a surrogate rover, and show that the rover effectively avoids dangers while making progress towards specified goals.

Spectral mapping, typically performed at the orbital scale, is hard at the rover scale as it necessitates larger coverage and field operable spectrometers. In this work, we propose using RGB cameras to assist spectral mapping. RGB cameras placed on a wide range of robotic platforms, including aerial vehicles, which can explore large regions compared to ground rovers. Our method uses a spectral model which learns the spatial relationship of spectra in a compressed feature space. We show that RGB data can contribute to this feature space and thereby enhance spectral reconstruction accuracy.

This work introduces range-based GP maps, which directly represent terrain by modeling the range from a LiDAR sensor as a Gaussian process (GP) in spherical space. Such a model aligns the predicted uncertainty from the GP regression with the uncertainty in the underlying sensor observations. Experimental evaluation on simulated natural terrain indicates that local range-based GP maps perform comparably to elevation-based methods when predicting terrain height, with the former producing more stable parameters and providing a better uncertainty representation. An aggregation method is proposed using the pose as an additional input to the GP. Unlike their elevation-based counterparts, range-based GP maps are capable of modeling overhangs and vertical obstacles with ease, demonstrated with examples of maps built on real-world data from a fully 3D subterranean environment.

Robotic explorers are essential tools for gathering information about regions that are inaccessible to humans. For applications like planetary exploration or search and rescue, robots use prior knowledge about the area to guide their search. Ergodic search methods find trajectories that effectively balance exploring unknown regions and exploiting prior information. In many search based problems, the robot must take into account multiple factors such as scientific information gain, risk, and energy, and update its belief about these dynamic objectives as they evolve over time. However, existing ergodic search methods either consider multiple static objectives or consider a single dynamic objective, but not multiple dynamic objectives. We address this gap in existing methods by presenting an algorithm called Dynamic Multi-Objective Ergodic Search (D-MO-ES) that efficiently plans an ergodic trajectory on multiple changing objectives. Our experiments show that our method requires up to nine times less compute time than a naïve approach with comparable coverage of each objective.