AI for Resilient Autonomous Operations and Scientific Discovery in Extreme and Unpredictable Environments

Introduction: Defining Extreme Environments and the Critical Role of AI-Driven Autonomous Systems.

• Extreme environments represent a class of settings that impose significant operational and technological challenges due to conditions far exceeding the normal parameters for human comfort and conventional technology. These environments are inherently resource-constrained, characterized by limitations in energy availability, computational power, communication bandwidth, and essential consumables. They are also fundamentally dynamic, exhibiting rapid and often unpredictable shifts in environmental conditions. Furthermore, these locations are often largely unknown, with sparse or incomplete prior data and understanding. Examples of such environments are diverse and include deep space missions navigating vacuum, radiation, and immense distances; sub-oceanic exploration contending with crushing pressure, darkness, and limited communication; polar research facing extreme cold, ice cover, and remote locations; active volcanic regions marked by high temperatures, toxic gases, and unpredictable eruptions; and post-disaster zones with collapsed infrastructure suffering from resource scarcity, hazardous conditions, and disrupted communication networks.

• Traditional human-operated systems encounter substantial limitations when attempting to function in these extreme conditions. The safety risks associated with human exposure to these hazardous environments are considerable. Logistical challenges in reaching and sustaining operations in remote and inaccessible locations are often complex and prohibitively costly. Significant communication delays, such as those experienced in deep space, can render real-time control impractical. Moreover, human endurance limitations, both physical and psychological, are severely tested under such demanding circumstances.

• Artificial intelligence (AI)-driven autonomous systems offer a paradigm shift in our ability to operate, make decisions, and conduct scientific investigations in these challenging settings. These systems hold the promise of sustained operation through robotic platforms powered by intelligent algorithms capable of functioning for extended periods without direct human intervention. They enable autonomous decision-making by processing vast amounts of sensor data and making real-time adjustments to plans and actions. Furthermore, AI can facilitate scientific discovery by autonomously analyzing data, generating novel hypotheses, and even guiding experimental processes in remote and inaccessible locations.

• This report will delve into the crucial research areas that are essential to unlock the full potential of AI in extreme environments. These areas encompass advanced perception in degraded and novel sensory conditions, on-board learning and adaptation to unforeseen circumstances, intelligent energy autonomy and resource management, explainable and verifiable autonomous decision-making, AI-driven hypothesis generation and autonomous experimentation for scientific discovery, and effective frameworks for human-autonomy teaming in remote exploration scenarios.

• Insight: The growing imperative to explore and conduct operations in extreme environments, fueled by scientific curiosity, the pursuit of resources, and critical security needs, necessitates a fundamental transition from predominantly human-centric approaches to autonomous systems that are intelligently driven by artificial intelligence. Human capabilities are inherently constrained by the severe conditions present in extreme environments. Factors such as safety, endurance, and the limitations of remote communication directly necessitate the development and deployment of autonomous systems. Artificial intelligence is not merely an enhancement but a foundational requirement for these systems to effectively perceive, reason, and act independently. This technological shift is not just about improving efficiency; it is about enabling access and operations in domains previously beyond sustained human reach.

Challenges in Extreme Environments: A Detailed Examination of the Operational Constraints.

• Resource Scarcity:

• Operating in extreme environments invariably involves confronting significant limitations in the availability of essential resources.

• Energy in these settings is often scarce, requiring reliance on distant power sources, potentially unreliable connections, or the development of in-situ energy generation capabilities.

• Computational power available on-board autonomous systems is frequently constrained by stringent size, weight, and power (SWaP) limitations, restricting the complexity of AI models that can be deployed.

• Communication bandwidth for transmitting data and receiving instructions from remote operators can be severely limited, especially in deep space or underwater environments.

• Available data for training AI models and understanding the environment may be sparse, particularly in largely unexplored regions, making it challenging to develop robust and reliable systems.

• Snippet ID: 125 underscores the substantial electricity demands associated with training and deploying large-scale generative AI models, such as OpenAI's GPT-4. The computational power needed for these models, often involving billions of parameters, can lead to staggering electricity consumption, putting pressure on power grids and increasing carbon dioxide emissions. This highlights the critical need for AI solutions designed for extreme environments to be exceptionally energy-efficient, as the power demands of conventional large AI models would be unsustainable in resource-constrained settings. Furthermore, the article notes the significant water requirements for cooling the hardware used in AI model training and deployment, which could also be a limiting factor in many extreme environments where water resources are scarce.

• Snippet ID: 128 explicitly identifies the considerable power resource challenges encountered when deploying onboard AI in space missions. The operation of AI systems in spacecraft necessitates substantial computational and power resources, and the need for high currents at low voltages presents a significant engineering hurdle in the space environment. This further emphasizes the fundamental constraint of power availability for AI in extreme conditions.

• Insight: The significant resource demands of traditional AI approaches present a major impediment to their effective deployment in extreme environments. This necessitates a fundamental shift towards the development of resource-aware AI, where algorithmic efficiency and optimized hardware utilization are paramount design considerations. The availability of resources in extreme environments directly dictates the operational envelope for autonomous systems. Limited energy restricts the complexity and runtime of AI models. Constrained bandwidth impacts the ability to communicate and process data remotely. Data scarcity hinders the development of robust AI through conventional training methods. Therefore, a concerted research effort is required to create AI solutions that can function effectively within these inherent resource limitations. This includes exploring novel AI architectures, efficient learning algorithms, and specialized hardware designed for low power consumption and high performance.

• Degraded Conditions:

• The operational effectiveness of autonomous systems in extreme environments is often significantly hampered by a variety of adverse environmental factors that degrade sensing and communication capabilities.

• Limited visibility is a common challenge in environments such as the deep sea, the polar regions during winter, and volcanic areas obscured by ash clouds, severely restricting the utility of optical sensors.

• Unreliable communication plagues sub-oceanic exploration due to the properties of water affecting radio waves, and deep space missions due to vast distances and signal attenuation. Similarly, post-disaster zones may experience widespread communication infrastructure damage.

• Autonomous systems may also encounter novel sensory inputs that their pre-trained AI models are not equipped to handle, such as unique acoustic signatures in the deep sea or non-visual cues essential for navigation in visually impaired conditions. Furthermore, in hostile environments, systems may face adversarial conditions where sensing inputs are intentionally manipulated to cause malfunctions or incorrect decisions.

• Snippet ID: 5 highlights that current autonomous driving algorithms, which heavily rely on the visible spectrum, suffer significant performance degradation in adverse weather conditions like fog, rain, and snow. While other spectral bands, such as near-infrared (NIR) and long-wave infrared (LWIR), can offer some enhancement, they lack the extensive datasets and benchmarks available for visible spectrum data. This underscores the broader challenge of achieving robust perception in degraded visual conditions, a problem prevalent across many extreme environments beyond just autonomous driving.

• Snippet ID: 12 further emphasizes this point by noting that the performance of existing object recognition algorithms can degrade substantially under challenging scenarios including rainy, foggy, and rainy night conditions. The presence of raindrops, light reflection, and low illumination pose significant obstacles to robust and accurate visual perception, highlighting the need for AI-driven solutions that can maintain performance despite these adverse factors.

• Insight: The inherent limitations of standard perception and communication technologies in extreme environments necessitate the development of AI systems that can maintain robust functionality despite degraded conditions. This requires innovations in advanced sensing techniques, intelligent sensor fusion strategies, and resilient signal processing algorithms. The fundamental ability of an autonomous system to operate effectively hinges on its capacity to accurately perceive its surroundings and reliably communicate with remote supervisors or other agents. Degraded conditions directly impair these critical functions. AI must play a central role in mitigating these impairments by enabling systems to extract meaningful information from noisy or obscured sensor data and to maintain connectivity despite challenging transmission environments. This involves not only enhancing the capabilities of individual sensors but also intelligently combining data from multiple sources to create a more complete and reliable understanding of the operational context.

• Unpredictability and Dynamicity:

• A defining characteristic of many extreme environments is their inherent unpredictability and rapid dynamism, posing significant challenges for autonomous systems designed to operate safely and effectively.

• Post-disaster zones are particularly volatile, with conditions on the ground shifting constantly due to factors such as aftershocks, the ongoing collapse of unstable structures, and the emergence of new hazards.

• Active volcanic regions are subject to sudden and dramatic changes, including unpredictable eruptions, lava flows, and the release of toxic gases, requiring systems to adapt swiftly to potentially catastrophic events.

• Even seemingly stable environments like deep space and sub-oceanic exploration can present unforeseen challenges, with encounters with previously unknown phenomena or sudden shifts in environmental parameters.

• Snippet ID: 129 underscores the crucial role of autonomy in highly uncertain scenarios, such as deploying robots on the Moon or Mars. In these environments, standard pre-programmed procedures may prove inadequate, and the substantial time delays in communication necessitate that robots possess the intelligence to analyze their surroundings and make informed decisions independently, without waiting for human input. This highlights the fundamental requirement for AI to enable autonomous decision-making in unpredictable settings.

• Snippet ID: 130 further elaborates on the limitations of AI in handling unpredictability, specifically in the context of autonomous vehicles. The article points out that while AI systems are trained on extensive datasets, no dataset can fully encompass the sheer diversity and unpredictability of real-world environments. As a result, autonomous systems may encounter novel situations not present in their training data, leading to potential failures in perception, decision-making, and control. This challenge of generalizing to unseen scenarios is equally pertinent to AI deployed in other extreme environments, where the range of potential conditions can be vast and difficult to anticipate.

• Insight: The dynamic and unpredictable nature of extreme environments demands that AI systems possess robust capabilities for real-time adaptation and continuous learning. These systems must be able to handle unforeseen circumstances, detect and respond to novel phenomena, and adjust their operational parameters in response to rapidly changing conditions. Autonomous systems operating in these highly variable settings cannot rely solely on pre-programmed responses or static models. AI must empower them to continuously monitor their environment, identify deviations from expected patterns, and dynamically adjust their behaviors to maintain safety and achieve mission objectives. This includes the ability to learn from new experiences, update their internal representations of the world, and make informed decisions in the absence of complete or certain information.

• Communication Limitations:

• Certain extreme environments impose significant constraints on the ability of autonomous systems to communicate effectively with remote human operators or other collaborating agents.

• Significant latencies, particularly pronounced in deep space missions, can result in delays of minutes or even hours in transmitting commands and receiving feedback, rendering real-time control impractical.

• Bandwidth limitations, common in remote terrestrial locations and underwater environments, restrict the amount of data that can be exchanged, potentially hindering the transmission of high-resolution sensor data or complex instructions.

• Intermittent connectivity can occur in various extreme environments due to factors such as weather conditions, terrain obstructions, or the operational constraints of communication relays, leading to periods where autonomous systems must operate without any external guidance.

• Snippet ID: 131 clearly illustrates this challenge, stating that due to the long communication distances involved in space exploration, robots on Mars cannot receive direct, immediate orders from Earth. This necessitates the integration of sophisticated AI systems that enable these robots to make decisions promptly at their location, without the prohibitive delays associated with waiting for instructions from mission control. This highlights the critical role of onboard autonomy in overcoming communication latency.

• Snippet ID: 129 further emphasizes the impact of communication limitations, noting that the time scales for communication between Earth and Mars are so long due to the vast distance that robot autonomy becomes even more crucial. Robots deployed on Mars must be capable of operating independently, without the possibility of relying on remote human intervention for every decision or action.

• Insight: The operational constraints imposed by communication limitations in extreme environments necessitate a high degree of onboard autonomy for AI-driven systems. These systems must be capable of making complex decisions, planning and executing tasks, and responding to their environment with minimal or no reliance on real-time human input. Furthermore, efficient strategies for utilizing the limited available bandwidth to transmit essential information, such as system status and critical scientific data, are of paramount importance. When autonomous systems are deployed in locations where communication with human operators is severely delayed, intermittent, or bandwidth-constrained, they must possess the intelligence to function effectively on their own. AI needs to provide the capacity for these systems to perceive, reason, and act independently, ensuring mission continuity and safety. Additionally, innovative approaches to data compression, prioritization of information, and potentially even distributed decision-making among multiple autonomous agents become critical for maximizing the value of the limited communication channels available.

Advanced Perception and Situational Awareness: AI Techniques for Robust Sensing, Sensor Fusion, and World Modeling in Degraded and Novel Sensory Conditions.

• Robust Sensing in Degraded Conditions:

• AI-powered techniques are essential for enhancing the capabilities of sensors operating in extreme environments where visibility is often limited. Image restoration algorithms, for example, can improve the quality of images captured in low-light or turbid conditions common in the deep sea.1 Noise filtering methods, facilitated by AI, can help to extract meaningful signals from noisy sensor data, improving the reliability of perception.2 Furthermore, leveraging non-visual sensors, such as infrared cameras for detecting heat signatures in polar environments 3, thermal sensors for volcanic activity monitoring, and short-wave infrared (SWIR) imaging for enhanced visibility in fog and smoke 5, can provide critical information when traditional visual sensing is impaired.

• Event-based cameras present a novel approach to visual sensing, offering significant advantages in situations with extreme lighting variations or rapid motion due to their high dynamic range and low latency.6 AI algorithms are specifically required to process the sparse, asynchronous event data produced by these sensors to extract relevant information about the environment and moving objects.8

• In underwater exploration, where visibility is often poor, AI plays a crucial role in enhancing sonar imaging. Techniques for noise reduction, advanced object detection algorithms for identifying submerged structures or marine life, and image segmentation methods for classifying different features on the seafloor are all enabled by AI.10

• Insight: The ability to extract meaningful information from sensor data that is inherently noisy, obscured, or derived from non-traditional modalities in extreme environments is critically dependent on the application of sophisticated AI-driven processing techniques. Standard sensing methods often fall short in the challenging conditions of extreme environments. AI's capacity to analyze and interpret sensor data, even when it is degraded or comes from unfamiliar sources, is fundamental to enabling autonomous systems to perceive their surroundings accurately and reliably. This involves not only cleaning up noisy data but also extracting high-level information and understanding from potentially ambiguous inputs.

• Sensor Fusion in Extreme Environments:

• A critical aspect of achieving comprehensive situational awareness in extreme environments is the intelligent fusion of data from multiple sensors. AI algorithms are essential for combining information from diverse sources such as visual cameras, LiDAR, radar, and sonar, each of which offers unique advantages and limitations depending on the environmental conditions.2 This synergistic approach allows autonomous systems to compensate for the weaknesses of individual sensors in specific situations, leading to a more robust and reliable understanding of the operational context.

• Sensor fusion can occur at various levels of abstraction, including raw data fusion, feature-level fusion, and decision-level fusion, and the choice of method often depends on the specific application and the nature of the sensors involved.2 AI algorithms must be capable of handling data that may be asynchronous, have different sampling rates, and potentially contain spatial or temporal misalignments to effectively integrate the information.15

• For instance, in the context of autonomous driving, which shares many challenges with operating in post-disaster zones, AI-enabled multi-sensor fusion aims to achieve more reliable perception under adverse weather conditions and even in cases of sensor degradation or failure.12 By intelligently combining the strengths of different sensors, such as the color information from cameras with the depth accuracy of LiDAR and the robustness of radar to weather, AI can create a more complete and accurate representation of the environment.

• Insight: AI-powered sensor fusion is a vital technique for creating a more complete, accurate, and resilient understanding of extreme environments by intelligently integrating data from a variety of sensing modalities, thereby overcoming the limitations inherent in relying on any single sensor.  Different sensors provide different types of information about the environment, and their performance can vary significantly depending on the conditions. AI acts as the intelligent orchestrator, taking the disparate data streams from these sensors and combining them in a way that maximizes the overall situational awareness of the autonomous system. This involves not just merging the data but also resolving inconsistencies, handling uncertainties, and extracting a unified and coherent understanding of the surroundings.

• World Modeling with Sparse and Unreliable Data:

• Creating and maintaining accurate world models in extreme environments is particularly challenging due to the often sparse and unreliable nature of the sensor data available. AI offers several techniques to address this issue, including methods like Sparse Data Diffusion (SDD) 17 and various sparse modeling approaches.18 These techniques are specifically designed to work effectively with datasets where a significant portion of the values may be missing or zero, which is common in situations with limited sensor coverage, communication dropouts, or inherent data sparsity.

• AI can also be employed for data imputation, where algorithms intelligently infer and reconstruct missing information based on the available data and learned patterns, leading to more complete and usable world models.14 In polar research, for example, AI has been successfully used to reconstruct missing historical climate data from limited observational records, demonstrating its ability to fill in gaps and provide a more comprehensive understanding of the environment over time.21

• Furthermore, AI can play a crucial role in handling the uncertainty associated with unreliable sensor data. By using probabilistic models and techniques for uncertainty estimation, AI can build world models that not only represent the environment but also quantify the level of confidence in that representation. This is vital for robust decision-making, as it allows autonomous systems to account for the inherent uncertainties when planning and executing actions in extreme environments.

• Insight: AI provides essential tools for constructing and maintaining coherent and informative representations of the world even when faced with the challenges of limited, noisy, and incomplete data that are characteristic of many extreme operating conditions. The ability of an autonomous system to effectively interact with its environment is predicated on having a sufficiently accurate and reliable model of that environment. In extreme settings where data is scarce and often of poor quality, AI's capacity to work with sparse information, infer missing details, and quantify uncertainty becomes paramount. This enables systems to navigate, plan, and make scientific inferences even when the sensory inputs are far from ideal.

• Handling Novel Sensory Inputs:

• To achieve a comprehensive understanding of extreme environments, autonomous systems often need to process and interpret data from non-traditional or novel sensors. Embodied AI, which focuses on developing intelligence within a physical form, emphasizes learning through interaction with the environment using a variety of senses.22 Architectures like the Predictive coding inspired, Variational Recurrent Neural Network (PV-RNN) are designed to integrate multiple sensory inputs simultaneously, including vision, proprioception, and even language instructions, enabling a more holistic understanding of the world.24

• Event cameras represent a significant departure from traditional frame-based imagers, providing asynchronous output based on changes in brightness.6 Interpreting this novel type of visual data requires specialized AI algorithms that can extract motion information and perform tasks like object tracking and SLAM.8

• Beyond visual and auditory inputs, robots can also leverage other sensory modalities, such as tactile sensing for understanding object properties, and even olfactory sensors (electronic noses) for detecting specific gases or substances in the environment.27 AI algorithms are needed to process and interpret these diverse non-visual data streams, allowing robots to gain a richer understanding of their surroundings.

• Insight: Autonomous systems operating in extreme environments must be equipped with AI that is flexible enough to process and learn from a wide array of sensory inputs, extending beyond traditional visual and auditory information to include novel modalities that can provide crucial insights into these challenging settings. The unique characteristics of extreme environments may necessitate the use of specialized sensors to gather relevant data. AI's ability to adapt to and interpret these novel sensory inputs is essential for enabling autonomous systems to fully perceive and interact with their surroundings. This includes developing new algorithms and potentially even biologically inspired approaches to process information from unfamiliar sensor types.

• Perception in Adversarial Conditions:

• In certain extreme environments, such as military operations or disaster response scenarios where malicious actors might be present, autonomous systems could face adversarial attempts to deceive their perception systems. AI plays a crucial role in enabling these systems to detect and mitigate such threats.30 Techniques like adversarial training aim to increase the robustness of AI models by exposing them to perturbed inputs during training, making them less susceptible to adversarial examples in real-world deployments.31

• Research in the field of autonomous driving, which is highly relevant to ground-based robots in post-disaster zones, has extensively explored the vulnerabilities of AI perception to adversarial attacks and the development of defense mechanisms.30 These findings can be leveraged for other autonomous systems operating in potentially hostile extreme environments.

• Sensor fusion offers an additional layer of defense against adversarial attacks, as inconsistencies between data from different sensors can potentially reveal attempts to manipulate the system's perception.13 AI algorithms can be designed to cross-validate sensor inputs and identify suspicious anomalies that might indicate an adversarial attack.

• Insight: For autonomous systems to operate safely and reliably in extreme environments where adversaries may be present, robust AI-driven mechanisms for detecting and defending against attacks on their perception systems are essential. The integrity of an autonomous system's perception is paramount for its safe and effective operation. In situations where adversaries might try to mislead the system through manipulated sensor inputs, AI must provide the tools to identify these attacks and maintain a correct understanding of the environment. This requires a proactive approach, anticipating potential threats and building resilience into the perception system.

On-board Learning, Adaptation, and Novelty Handling: AI Architectures for Continuous Learning, Model Adaptation, Self-Diagnosis, and Response to Unforeseen Phenomena.

• Continuous Learning from Sparse Data:

• In the context of extreme environments, where data acquisition can be infrequent or limited, AI architectures must support continuous learning from sparse data streams. Lifelong learning algorithms are designed to enable AI systems to acquire new knowledge sequentially from multiple tasks without suffering from catastrophic forgetting, a critical capability for long-duration missions or operations in evolving environments.35

• Sparse and Expandable Networks (SEN) offer a promising approach by allowing AI models to handle multiple tasks concurrently while maintaining sparsity, which is particularly beneficial for resource-constrained devices that may only have access to limited data for each task.20

• Transfer learning techniques can also be highly valuable, allowing models pre-trained on large datasets (potentially from different but related domains) to be adapted to new tasks or environments with significantly less data, which is often the case when deploying AI in novel extreme settings.37

• Insight: The inherent scarcity of data in many extreme environments necessitates the development and deployment of AI systems capable of continuous learning and adaptation, allowing them to improve their performance and acquire new knowledge incrementally from limited experiences. Unlike many traditional machine learning applications where large, labeled datasets are available, AI in extreme environments must often learn on the fly from small amounts of data. This requires algorithms that can efficiently extract information from sparse data, retain previously learned knowledge, and adapt to new situations as more data becomes available over time.

• Model Adaptation to Environmental Shifts:

• Extreme environments are often characterized by dynamic conditions and unforeseen shifts. AI models deployed in these settings must possess the ability to adapt their parameters and behaviors in response to these changes. Adaptive AI approaches enable robotic systems to perform reliably and efficiently across diverse and challenging conditions by continuously adjusting their operation based on real-time environmental feedback.38

• Reinforcement learning (RL) provides a powerful paradigm for training intelligent agents that can handle unknown and extreme conditions by learning through interaction with their environment and adapting their policies based on the rewards they receive.39 This is particularly useful in situations where the environmental dynamics are complex or difficult to model explicitly.

• Techniques such as domain adaptation allow AI models trained in one set of environmental conditions (e.g., a simulation or a less extreme environment) to be effectively applied to a different but related domain, reducing the need for extensive retraining when the environment shifts.14

• Insight: The dynamic nature of extreme environments requires AI models to be flexible and capable of adapting to unforeseen changes, ensuring that autonomous systems can maintain optimal performance and safety despite environmental variability. Pre-trained AI models may not always be perfectly suited to the specific conditions encountered in a particular extreme environment, and these conditions can change over the course of a mission. Therefore, the ability for the AI to monitor the environment, detect changes, and adjust its internal models and control strategies accordingly is crucial for long-term success.

• Novelty Detection and Handling:

• Operating in largely unknown extreme environments means that autonomous systems will inevitably encounter novel phenomena and unexpected situations that were not anticipated during their design or training. AI systems must be equipped with robust novelty detection capabilities to identify these unfamiliar inputs and trigger appropriate responses. Various novelty detection algorithms exist, including those based on autoencoders, which have shown promise in detecting unusual features in planetary exploration imagery.40

• Beyond simply detecting novelties, AI systems also need mechanisms to handle them effectively. This might involve triggering a pre-programmed safe behavior, alerting a remote human operator if communication is available, or initiating an autonomous exploration or analysis of the novel phenomenon to understand it better.41 In some cases, the system might even need to learn a new model or behavior to appropriately interact with the novelty in the future.

• AI can also employ intrinsic motivation, where the system is inherently driven to seek out and explore novel aspects of its environment, which can be particularly valuable in scientific discovery tasks in unexplored extreme environments.46

• Insight: The ability to identify and appropriately manage novel stimuli and events is paramount for autonomous systems to operate safely and effectively in the inherently uncertain and unexplored territories of extreme environments. Because extreme environments are often poorly understood, autonomous systems cannot be prepared for every possible scenario. AI must provide the capacity to recognize something new or unexpected and to react in a way that ensures safety and potentially contributes to scientific understanding of the novel phenomenon.

• Self-Diagnosis and Recovery from Partial System Failures:

• For autonomous systems to achieve sustained operation in remote and potentially hazardous extreme environments, they must possess the capability for self-diagnosis and recovery from partial system failures. Self-healing AI systems are designed to continuously monitor their own performance, detect anomalies or faults in their hardware or software, and autonomously initiate corrective actions to restore functionality without human intervention.47

• These self-healing mechanisms can involve a range of techniques, including anomaly detection algorithms to identify deviations from normal behavior, root cause analysis to pinpoint the source of the problem, fault isolation to prevent cascading failures, and automated repair or reconfiguration procedures to restore the affected system to an operational state.48 In space applications, for instance, research has explored runtime self-test mechanisms and dynamic partial reconfiguration as methods for rapidly detecting and correcting errors in AI accelerators.51

• The ability for autonomous systems to self-diagnose and recover from failures is critical in extreme environments where human intervention for maintenance or repairs may be significantly delayed or even impossible. This capability enhances the overall resilience and mission success rate of these systems.

• Insight: The sustained and reliable operation of autonomous systems in extreme environments hinges on their ability to autonomously identify and resolve internal failures, minimizing downtime and ensuring mission continuity in the absence of timely human assistance. The harsh conditions and demanding operational profiles of extreme environments increase the likelihood of system malfunctions. AI-driven self-diagnosis and recovery capabilities are essential for enabling these systems to maintain functionality and safety over extended periods, without relying on immediate support from remote human operators.

Energy Autonomy and Resource Management: AI Strategies for Long-Term Energy Management, Efficient Power Utilization, and Potential In-Situ Resource Utilization.

• Long-Term Energy Management:

• For autonomous systems to undertake extended missions in extreme environments without human intervention, intelligent energy management is paramount. AI can play a crucial role by predicting future energy needs based on mission plans, environmental conditions, and historical usage patterns, allowing for proactive allocation of power resources.53 Autonomous agents can dynamically balance the load demand between different subsystems and optimize the charging and discharging cycles of onboard energy storage systems to maximize operational time.54 In robotic platforms, AI algorithms can plan energy-efficient paths and strategically manage the power consumption of various components, such as sensors and actuators, to extend the overall mission duration.55

• Insight: Intelligent energy management driven by AI is a fundamental requirement for enabling long-duration autonomous missions in extreme environments where access to external power sources or opportunities for recharging are severely limited or non-existent. The total energy available to an autonomous system in an extreme environment is finite. AI must make sophisticated decisions about how this energy is used over the entire course of the mission, ensuring that critical functions are prioritized and that power is conserved whenever possible to extend the operational lifespan of the system.

• Efficient Power Utilization:

• Minimizing energy consumption is a critical design consideration for AI deployed in resource-constrained extreme environments. Embedded AI, which involves running AI models directly on low-power hardware at the edge, offers a solution by enabling real-time processing of sensor data without the need for energy-intensive cloud computing.56 AI algorithms themselves must be designed for efficiency, utilizing techniques like model compression and optimization to reduce their computational footprint and power requirements.56 Furthermore, AI can contribute to efficient power utilization by optimizing the operation of the entire autonomous system, for example, by planning smoother and more direct movement trajectories for robots.59

• Insight: The tight power budgets typical of autonomous systems operating in extreme environments necessitate a strong focus on energy efficiency at all levels of AI design and implementation, from the choice of algorithms to the optimization of hardware. Every watt of energy saved by the AI and the autonomous system as a whole translates directly into increased operational time or the ability to carry more scientific payload. Therefore, developing and deploying highly energy-efficient AI is not just a desirable feature but a crucial factor in the success of missions in these challenging settings.

• Potential In-Situ Resource Utilization (ISRU):

• For long-term human presence and sustained robotic operations in extraterrestrial environments, the ability to utilize locally available resources is essential. AI can play a pivotal role in enabling In-Situ Resource Utilization (ISRU) by analyzing sensor data from orbiters and landers to identify potential deposits of valuable resources, such as water ice on the Moon or oxygen-rich regolith on Mars.60 Autonomous robots, guided by AI, can then be tasked with the extraction and processing of these raw materials to produce usable resources like water, breathable air, or even rocket propellant.60 AI-driven systems can manage the complex processes involved in refining these resources, optimizing energy consumption and maximizing yield.60

• Insight: AI is a critical enabler for realizing the transformative potential of ISRU, allowing autonomous systems to become more self-sufficient in extraterrestrial environments, reducing their reliance on costly and logistically challenging resupply missions from Earth. The ability to live off the land, so to speak, is fundamental for establishing a long-term presence beyond Earth. AI provides the intelligence necessary for autonomous systems to locate, extract, and process resources found in space, paving the way for extended human exploration and the potential for off-world settlements.

Explainable and Verifiable Autonomous Decision-Making: Methods for Ensuring Transparency, Trustworthiness, and Ethical Alignment in AI Decisions, with Considerations for Human Oversight in Remote Scenarios.

• Explainable AI (XAI) for Autonomous Systems:

• In safety-critical extreme environments, it is paramount that the decision-making processes of autonomous AI systems are transparent and understandable to human operators. Explainable AI (XAI) aims to address the "black box" nature of many complex AI models by providing insights into how they arrive at their conclusions. Hybrid AI approaches that combine machine learning with symbolic reasoning can offer more interpretable explanations for robot actions, such as in navigation.63 Techniques like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) can be employed to understand the feature importance and local behavior of complex models used in critical applications like robotic power management.64 Building trust in AI, especially in high-stakes remote operations, hinges on the ability of human supervisors to understand the rationale behind the autonomous system's decisions.65

• Insight: Transparency in the decision-making of autonomous systems is not merely a desirable feature but a fundamental requirement for fostering user trust and enabling effective human-autonomy teaming, particularly in the high-risk contexts of extreme environments. When human operators are responsible for overseeing autonomous systems in remote and potentially dangerous settings, they must have a clear understanding of why the AI is making certain choices. This transparency allows them to identify potential errors, biases, or unforeseen consequences and to intervene appropriately when necessary. Without explainability, it is difficult for humans to trust and rely on the autonomous system's judgments, potentially leading to disuse or, conversely, to over-reliance without a full understanding of the system's capabilities and limitations.

• Verifiable Autonomous Decision-Making:

• For autonomous systems operating in safety-critical extreme environments, ensuring the correctness and safety of their decision-making algorithms is of utmost importance. Formal verification methods provide a rigorous mathematical approach to proving that a system's behavior adheres to a given specification, offering a higher level of assurance than traditional testing-based methods.68 These techniques can be used to verify properties such as safety constraints, liveness conditions, and real-time performance requirements. Runtime monitoring offers a complementary approach by continuously checking the system's behavior against predefined safety specifications during operation.75 Furthermore, self-certification mechanisms can enable autonomous systems to dynamically assess whether they are operating within their certified safety boundaries.68

• Insight: The application of formal verification and runtime monitoring techniques is essential for establishing the trustworthiness of autonomous decision-making in extreme environments, especially in scenarios where system failures could lead to catastrophic consequences. In high-risk domains like space exploration, deep-sea operations, and disaster response, the potential for autonomous systems to make incorrect or unsafe decisions is unacceptable. Formal verification provides a mathematical guarantee that the system will behave as intended under all possible conditions, while runtime monitoring acts as a continuous safety net, detecting and potentially mitigating any deviations from safe operation. These methods are crucial for building confidence in the deployment of autonomous systems in these challenging settings.

• Ethical Alignment of Autonomous Decisions:

• As autonomous AI systems become more prevalent in extreme environments, it is crucial to ensure that their decisions are not only effective and safe but also aligned with mission objectives and ethical guidelines. This requires the integration of ethical frameworks into the design of AI algorithms.79 Leaders and developers must consider the broader societal impacts of autonomous systems and strive to uphold core values such as integrity, fairness, and accountability in their design and deployment.81 This includes addressing potential biases in AI algorithms that could lead to discriminatory or unintended outcomes.82 Furthermore, the principles of transparency and responsibility are paramount in ensuring the ethical use of AI in autonomous systems.79

• Insight: The ethical implications of autonomous decision-making in extreme environments must be proactively addressed to ensure responsible innovation and to maintain public trust in these powerful technologies. Autonomous systems operating in remote and high-stakes settings will face complex situations that may involve ethical considerations. It is essential to design these systems with a clear understanding of the values and principles that should guide their actions, ensuring that they operate in a way that is consistent with human ethical norms and the overall goals of the mission. This requires careful consideration of potential biases in the data and algorithms, as well as the development of mechanisms for accountability and oversight.

• Human Oversight and Intervention in Time-Delayed Remote Scenarios:

• In many extreme environments, particularly those involving remote exploration, significant communication latencies can hinder real-time human intervention in the operations of autonomous systems. Optimal frameworks for human-autonomy teaming must therefore account for these delays. Supervisory control models allow human operators to provide high-level goals and monitor the system's progress, intervening only when necessary.84 Human-in-the-Loop (HITL) systems integrate human expertise at critical decision points, enabling operators to guide the autonomous system when it encounters complex or ambiguous situations.85 Adjustable autonomy approaches provide the flexibility to dynamically vary the level of autonomy of the AI agent, allowing for a seamless transition between autonomous operation and human control as the situation demands.87 User interfaces for remote operation must be carefully designed to provide sufficient situational awareness to human operators despite the time delays and bandwidth limitations.90

• Insight: Effective human oversight of autonomous systems in remote, time-delayed scenarios necessitates the implementation of adjustable autonomy frameworks and user interfaces that enable operators to maintain awareness and intervene strategically, despite the challenges posed by communication limitations. While the goal is to enable autonomous systems to operate independently in extreme environments, the need for human intervention in certain circumstances remains. The challenge lies in designing systems that can effectively support this oversight, even when there are significant delays in communication. This requires careful consideration of how to present information to the human operator, how to allow for timely and effective intervention, and how to manage the transitions between autonomous and human-controlled operation.

AI-Driven Hypothesis Generation and Autonomous Experimentation: Designing AI Systems for Autonomous Scientific Inquiry and Knowledge Discovery in Remote Settings.

• Autonomous Hypothesis Generation:

• Artificial intelligence, particularly through multi-agent systems leveraging large language models (LLMs), offers the potential to revolutionize scientific discovery in remote environments by autonomously generating research hypotheses based on the analysis of vast amounts of incoming data and existing scientific knowledge.92 These AI systems can identify subtle patterns, unexplored connections, and emerging trends in complex datasets that might be overlooked by human researchers working in isolation.92 Frameworks like SciAgents utilize knowledge graphs to establish relationships between diverse scientific concepts, enabling the AI to formulate evidence-driven hypotheses that align with unmet research needs.99

• Insight: The ability of AI to autonomously formulate scientific hypotheses in remote settings can significantly accelerate the pace of discovery by providing novel research directions and insights that can then be further investigated. In extreme environments, where human scientists may have limited access or the volume of data collected is overwhelming, AI can serve as a powerful tool for sifting through information, identifying potential areas of interest, and proposing testable hypotheses that can guide further scientific inquiry. This capability is particularly valuable in exploring the largely unknown aspects of these environments.

• Autonomous Experiment Design and Execution:

• Beyond generating hypotheses, AI systems can also be designed to autonomously plan and execute experiments to test these hypotheses in remote environments. AI-driven experimental design frameworks can identify the most relevant variables, optimize experimental parameters, and propose efficient strategies for data acquisition.92 Self-driving laboratories (SDLs), which integrate AI with robotic instrumentation, represent a significant step towards fully autonomous scientific experimentation, capable of carrying out complex protocols without human intervention.103 These systems can manage the entire experimental workflow, from hypothesis generation to execution and data analysis, potentially leading to a significant acceleration of the scientific process.103

• Insight: AI's capacity to automate the design and execution of scientific experiments in remote settings can overcome logistical challenges and enable continuous, unsupervised research, leading to more rapid advancements in our understanding of these environments. Conducting experiments in extreme environments can be difficult, dangerous, or resource-intensive for humans. AI-controlled robotic systems can perform these tasks autonomously, allowing for the systematic testing of hypotheses and the collection of valuable data even in the most challenging locations. This capability is particularly important for long-duration missions or continuous monitoring efforts.

• Autonomous Iteration on Scientific Understanding:

• A key aspect of the scientific method is the iterative process of refining hypotheses based on experimental results. AI systems can be designed to autonomously analyze the data collected from their own experiments, identify patterns and anomalies, and use this feedback to revise their hypotheses and plan subsequent investigations.103 Unsupervised AI models can even discover novel relationships and insights within the data that might not be immediately apparent to human researchers, leading to a continuous cycle of learning and knowledge discovery.109 This capability for autonomous iteration is particularly valuable in remote extreme environments where communication delays or bandwidth limitations might hinder frequent interaction with human scientists.

• Insight: AI's ability to autonomously analyze experimental outcomes and refine its understanding through iterative experimentation in remote settings enables a more efficient and potentially more profound advancement of scientific knowledge about these challenging environments. The scientific process is inherently iterative. AI can automate this iteration by not only generating and testing hypotheses but also by learning from the results and using that learning to guide future research. This closed-loop approach allows for a more rapid and potentially unbiased exploration of the scientific landscape in extreme environments.

Human-Autonomy Teaming for Remote Exploration: Optimal Frameworks for Collaboration with Significant Communication Latencies and Bandwidth Limitations, Focusing on Shared Situational Awareness and Adjustable Autonomy.

• Frameworks for Human-Autonomy Teaming (HAT):

• In the context of remote exploration in extreme environments, effective collaboration between human supervisors and autonomous AI agents is crucial. Human-Autonomy Teaming (HAT) frameworks emphasize a synergistic partnership where humans and AI work interdependently towards shared mission goals.111 These frameworks recognize the unique strengths of both humans (e.g., high-level reasoning, intuition, adaptability to truly novel situations) and AI (e.g., processing large datasets, rapid execution of tasks, sustained operation in harsh conditions). Key elements of successful HAT include establishing clear roles and responsibilities, facilitating effective communication, and fostering trust between human and AI team members.116

• Insight: Optimal remote exploration outcomes are likely to be achieved through well-designed human-autonomy teaming frameworks that strategically leverage the complementary capabilities of both human operators and intelligent AI agents. Neither humans nor fully autonomous AI systems are likely to be perfectly effective in the complex and uncertain conditions of remote extreme environments. HAT provides a way to combine the strengths of both, with humans providing high-level direction and oversight, and AI handling the more routine or computationally intensive tasks. The key is to define how this collaboration will work most effectively.

• Shared Situational Awareness in Remote Teams:

• Maintaining a shared understanding of the environment, the autonomous system's state, and the mission objectives is vital for effective human-autonomy teaming, especially when dealing with significant communication latencies and bandwidth limitations in remote exploration scenarios.90 User interfaces must be carefully designed to provide human operators with sufficient and timely information about the autonomous agent's perceptions, plans, and actions, despite the communication constraints.116 Conversely, the autonomous agent may also need to convey its understanding and intentions to the human supervisor in a clear and concise manner. Techniques such as providing explanations of AI decisions and visualizing the robot's internal state can contribute to shared situational awareness.116

• Insight: Enabling both human operators and remote AI agents to maintain a common and accurate understanding of the operational context, despite communication challenges, is essential for effective coordination and decision-making in collaborative exploration efforts. When humans and AI are working together remotely, it is crucial that both parties have a consistent and up-to-date picture of what is happening. Communication delays and limited bandwidth can make this challenging, so innovative approaches to information sharing and visualization are needed to ensure that both the human and the AI are on the same page.

• Adjustable Autonomy in Remote Exploration:

• The concept of adjustable autonomy, where the level of control and decision-making authority can be dynamically shifted between the human operator and the autonomous AI agent, is particularly relevant for remote exploration scenarios characterized by communication challenges.87 This flexibility allows the AI to handle routine tasks and react quickly to immediate situations, while also enabling human operators to step in and take control when faced with complex, novel, or safety-critical events that the AI may not be equipped to handle. The ability to adjust the level of autonomy can also help to optimize workload and maintain operator engagement over long-duration missions. Risk-based approaches to adjustable autonomy can help guide the system in determining when human intervention is most needed.88

• Insight: Providing the capability to dynamically adjust the level of autonomy in remote exploration missions allows for an optimal balance between the efficiency of autonomous operation and the critical oversight and intervention of human expertise, especially in the face of communication limitations. The ideal level of autonomy for a remote exploration task may vary depending on the specific circumstances. Adjustable autonomy provides the flexibility to adapt to these changing needs, allowing the AI to operate independently when appropriate and seamlessly transition control to a human operator when their judgment or intervention is required. This ensures that the strengths of both humans and AI are effectively utilized throughout the mission.

• Trust Calibration Between Human Supervisors and Remote AI Agents:

• For effective human-autonomy teaming in remote exploration, it is essential to establish and maintain an appropriate level of trust between the human supervisors and the autonomous AI agents. Trust calibration involves ensuring that the human operator's trust in the AI is aligned with the AI's actual capabilities and reliability.120 Over-trust can lead to complacency and a failure to intervene when necessary, while under-trust can result in the inefficient disuse of the AI's capabilities and increased workload for the human operator. Adaptive trust calibration methods can monitor the operator's reliance behavior and provide cues to encourage a more accurate assessment of the AI's trustworthiness.120 Transparency in the AI's decision-making processes also plays a crucial role in building and maintaining appropriate levels of trust.122

• Insight: Achieving and maintaining a well-calibrated level of trust between human supervisors and remote AI agents is critical for the success of collaborative exploration missions, ensuring that human operators can effectively leverage the AI's capabilities while remaining vigilant for potential errors or limitations. The effectiveness of human-autonomy teams depends not only on the technical capabilities of the AI but also on the human operator's willingness to rely on it appropriately. Trust calibration ensures that this reliance is neither excessive nor insufficient, maximizing the overall performance and safety of the team. This requires a continuous process of communication, feedback, and potentially even interventions to help humans develop an accurate understanding of the AI's strengths and weaknesses.

Cross-Environmental Synergies and Lessons Learned: Identifying Common AI Approaches and Best Practices Applicable Across Diverse Extreme Environments.

• Despite the distinct challenges posed by each extreme environment, there are several common AI approaches and best practices that demonstrate applicability across these diverse settings. Robust perception in degraded conditions is a universal need, addressed by AI-enhanced sensing, sensor fusion, and novel sensor modalities. On-board learning from sparse data is crucial in resource-constrained scenarios, regardless of whether they are in space, underwater, or in remote terrestrial locations. Energy efficiency is a paramount concern for all long-duration autonomous missions in extreme environments, driving the need for optimized AI algorithms and hardware. Finally, the importance of explainable and verifiable AI for ensuring safety and building trust transcends the specific environment.

• Successful AI techniques like deep learning for feature extraction and pattern recognition, reinforcement learning for adaptive control, and various methods for novelty detection and handling have found utility in multiple extreme environment domains. Best practices include prioritizing modular and adaptable AI architectures that can be tailored to specific mission requirements, focusing on developing energy-efficient solutions from the outset, and emphasizing the importance of building and maintaining user trust through transparency and reliability.

• Insight: Recognizing and leveraging the commonalities in challenges and successful AI solutions across different extreme environments can significantly accelerate the development and deployment of autonomous systems for exploration and operation in these diverse and demanding settings. While the specific environmental conditions may vary greatly, the fundamental requirements for autonomous AI often overlap. By identifying these common threads, researchers and engineers can avoid reinventing the wheel and instead adapt and build upon existing knowledge and technologies developed for other extreme environments. This cross-pollination of ideas and techniques can lead to more rapid progress and more robust solutions.

Environmental and Sustainability Considerations for AI in Extreme Environments: Analyzing the Resource Consumption and Ecological Impact of AI Deployments.

• The increasing reliance on AI for operations in extreme environments necessitates a careful consideration of the technology's own environmental footprint. The training of complex AI models, particularly large language models, demands significant computational resources, leading to substantial energy consumption and associated carbon emissions.125 Furthermore, the data centers that house these AI workloads often require considerable amounts of water for cooling, which can strain local ecosystems.125 The manufacturing of the specialized hardware used in AI systems, including GPUs and other processors, relies on critical minerals and rare earth elements, the extraction of which can have detrimental environmental consequences.126

• In ecologically sensitive extreme environments, such as polar regions and the deep sea, the deployment of AI-powered autonomous systems must be carefully managed to minimize disturbance to fragile ecosystems.127 This includes considering the potential impacts of noise pollution from underwater robots, the physical footprint of research infrastructure, and the risk of introducing contaminants.

• Strategies for developing more sustainable AI solutions for extreme environments include optimizing AI algorithms for energy efficiency to reduce power consumption, exploring the use of renewable energy sources to power AI infrastructure, and designing hardware with a focus on minimizing the use of environmentally harmful materials. Furthermore, in ecologically sensitive areas, AI can be used to monitor and mitigate the environmental impact of autonomous operations.

• Insight: While AI offers immense potential for advancing our understanding and utilization of extreme environments, it is crucial to address its own environmental and sustainability implications, particularly in ecologically fragile areas and in light of the resource constraints often inherent in these settings. The very environments we seek to explore and protect through AI-driven autonomous systems are often highly sensitive to environmental change and resource depletion. It is therefore essential that the AI solutions we develop and deploy are themselves sustainable and do not inadvertently contribute to the problems we are trying to solve. This requires a holistic approach that considers the entire lifecycle of AI systems, from the sourcing of materials to their energy consumption and ultimate disposal.

Conclusion and Future Research Directions: Summarizing Key Findings and Proposing Avenues for Further Investigation and Development.

• Artificial intelligence stands as a pivotal enabling technology for achieving resilient autonomous operations and groundbreaking scientific discovery across a diverse range of extreme environments.

• Despite the significant advancements highlighted in this report, substantial challenges persist in areas such as resource scarcity, the complexities of perception in degraded conditions, the inherent unpredictability of these environments, and the limitations of communication in remote settings.

• Future research endeavors should prioritize:

• The development of highly energy-efficient AI algorithms and specialized hardware tailored for the stringent power constraints of autonomous systems in extreme environments.

• Innovations in robust perception techniques capable of functioning reliably across a spectrum of adverse conditions and utilizing novel sensory inputs.

• Advancements in on-board learning and adaptation mechanisms that enable autonomous systems to continuously improve and respond effectively to dynamic and unforeseen environmental shifts.

• The creation of explainable and verifiable AI frameworks to ensure the transparency, trustworthiness, and ethical alignment of autonomous decision-making, particularly in safety-critical applications.

• The further exploration and refinement of AI-driven methodologies for autonomous scientific inquiry, including hypothesis generation, experiment design and execution, and iterative knowledge discovery in remote locations.

• The optimization of human-autonomy teaming frameworks that facilitate seamless and effective collaboration between human supervisors and remote AI agents, accounting for communication latencies and bandwidth limitations.

• A thorough investigation into the environmental and sustainability implications of AI deployments in extreme environments, with a focus on minimizing resource consumption and ecological impact.

• To accelerate progress in this critical field, fostering interdisciplinary collaborations among AI researchers, roboticists, and domain experts in space exploration, oceanography, polar science, volcanology, and disaster response is essential.

• The development of standardized benchmark datasets that accurately reflect the unique challenges of extreme environments, along with the creation of high-fidelity simulation environments, will be crucial for evaluating and comparing the performance of different AI approaches.

• Finally, increased emphasis must be placed on rigorous field testing and validation of AI-driven autonomous systems in relevant extreme environments to ensure that research findings translate into practical and reliable solutions for real-world applications.

Works cited

1. Blue w(AI)ve: SeaDeep Case Study - Gulf of Maine Research Institute, accessed on May 18, 2025, https://gmri.org/stories/blue-waive-seadeep-case-study/

2. Sensor Fusion and Its Role In Rugged Edge AI - Premio Inc, accessed on May 18, 2025, https://premioinc.com/blogs/blog/sensor-fusion-and-its-role-in-rugged-edge-ai

3. A new dataset of Arctic images will spur artificial intelligence research - MIT News, accessed on May 18, 2025, https://news.mit.edu/2023/new-dataset-arctic-images-will-spur-artificial-intelligence-research-0724

4. Developing Artificial Intelligence To Find Ice Seals And Polar Bears From The Sky, accessed on May 18, 2025, https://www.fisheries.noaa.gov/feature-story/developing-artificial-intelligence-find-ice-seals-and-polar-bears-sky

5. [2504.07603] RASMD: RGB And SWIR Multispectral Driving Dataset for Robust Perception in Adverse Conditions - arXiv, accessed on May 18, 2025, https://arxiv.org/abs/2504.07603

6. Event-based Vision Sensor (EVS) Technology | Image Sensor for ..., accessed on May 18, 2025, https://www.sony-semicon.com/en/technology/industry/evs.html

7. Event camera - Wikipedia, accessed on May 18, 2025, https://en.wikipedia.org/wiki/Event_camera

8. Event-based Vision, Event Cameras, Event Camera SLAM, accessed on May 18, 2025, https://rpg.ifi.uzh.ch/research_dvs.html

9. Event-based Neuromorphic Perception and Computation: The Future of Sensing and AI - Edge AI and Vision Alliance, accessed on May 18, 2025, https://www.edge-ai-vision.com/wp-content/uploads/2022/05/T1T02_Benosman_CMU_2022.pdf

10. JMSE | Special Issue : Artificial Intelligence Applications in ... - MDPI, accessed on May 18, 2025, https://www.mdpi.com/journal/jmse/special_issues/E0G55Z87RF

11. Autonomous Sonar Object Detection Underwater - IoT Use Case, accessed on May 18, 2025, https://iotusecase.com/en/solution-examples/real-time-sonar-object-recognition-for-an-autonomous-examination-robot/

12. Robust Perception Under Adverse Conditions for Autonomous Driving Based on Data Augmentation | Request PDF - ResearchGate, accessed on May 18, 2025, https://www.researchgate.net/publication/374821822_Robust_Perception_Under_Adverse_Conditions_for_Autonomous_Driving_Based_on_Data_Augmentation

13. AI in Autonomous Vehicles: Sensor Fusion, Path Planning, and Safety Challenges, accessed on May 18, 2025, https://www.researchgate.net/publication/390233508_AI_in_Autonomous_Vehicles_Sensor_Fusion_Path_Planning_and_Safety_Challenges

14. How AI Enhances LiDAR Data for Object Recognition, Classification, accessed on May 18, 2025, https://govcomm.us/ai-powered-lidar/

15. Sensors Fusion: An Overview, accessed on May 18, 2025, https://www.azosensors.com/article.aspx?ArticleID=3123

16. (PDF) Benchmarking Robustness of AI-Enabled Multi-sensor Fusion Systems: Challenges and Opportunities - ResearchGate, accessed on May 18, 2025, https://www.researchgate.net/publication/371346882_Benchmarking_Robustness_of_AI-enabled_Multi-sensor_Fusion_Systems_Challenges_and_Opportunities

17. Sparse Data Generation Using Diffusion Models - arXiv, accessed on May 18, 2025, https://arxiv.org/html/2502.02448v1

18. Using Sparse Modeling in AI: A Human Centric, Explainable Approach - HACARUS INC., accessed on May 18, 2025, https://hacarus.com/tech/using-sparse-modeling-in-ai-a-human-centric-explainable-approach/

19. How does sparse modeling work? - HACARUS INC., accessed on May 18, 2025, https://hacarus.com/thoughts-ai/20221016/

20. Sparse and Expandable Network for Google's Pathways - Frontiers, accessed on May 18, 2025, https://www.frontiersin.org/journals/big-data/articles/10.3389/fdata.2024.1348030/full

21. How artificial intelligence can help us understand the Earth better | UiT, accessed on May 18, 2025, https://en.uit.no/nyheter/forskerhjornet/845542/how_artificial_intelligence_can_help_us_understan

22. What is Embodied AI? A Guide to AI in Robotics - Encord, accessed on May 18, 2025, https://encord.com/blog/embodied-ai/

23. New, embodied AI reveals how robots and toddlers learn to understand - EurekAlert!, accessed on May 18, 2025, https://www.eurekalert.org/news-releases/1071383

24. New, embodied AI reveals how robots and toddlers learn to understand, accessed on May 18, 2025, https://www.oist.jp/news-center/news/2025/1/23/new-embodied-ai-reveals-how-robots-and-toddlers-learn-understand

25. AI Model Exhibits Child-like Generalization Abilities - AZoRobotics, accessed on May 18, 2025, https://www.azorobotics.com/News.aspx?newsID=15656

26. Revolutionary Embodied AI Sheds Light on How Robots and Toddlers Learn to Understand the World - BIOENGINEER.ORG, accessed on May 18, 2025, https://bioengineer.org/revolutionary-embodied-ai-sheds-light-on-how-robots-and-toddlers-learn-to-understand-the-world/

27. Robotics: Five Senses plus One—An Overview - MDPI, accessed on May 18, 2025, https://www.mdpi.com/2218-6581/12/3/68

28. The Journey Towards Human-Like Senses in Robots, accessed on May 18, 2025, https://www.bitsathy.ac.in/blog/the-journey-towards-human-like-senses-in-robots/

29. Transferring Implicit Knowledge of Non-Visual Object Properties across Heterogeneous Robot Morphologies - Tufts University, accessed on May 18, 2025, https://www.eecs.tufts.edu/~gtatiya/files/2023/IEEE_ICRA_2023_gtatiya.pdf

30. Adversarial Examples in Environment Perception for Automated Driving - arXiv, accessed on May 18, 2025, https://arxiv.org/html/2504.08414v1

31. How to increase robustness of AI perception for safety-critical domains - Bosch Global, accessed on May 18, 2025, https://www.bosch.com/stories/how-to-increase-robustness-of-ai-perception/

32. Adversarial Attacks on Deep Learning-Based Perception in ..., accessed on May 18, 2025, https://www.usi.ch/en/feeds/31244

33. Robust AI based perception and guidance for autonomous vehicles - City Research Online, accessed on May 18, 2025, https://openaccess.city.ac.uk/id/eprint/34241/

34. Security of AI-enabled Perception Systems in Autonomous Driving, accessed on May 18, 2025, https://www.cs.fsu.edu/talk20240227/

35. Advancing autonomy through lifelong learning: a survey of autonomous intelligent systems - PMC - PubMed Central, accessed on May 18, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC11027131/

36. AFRL-RI-RS-TR-2022-146 - DTIC, accessed on May 18, 2025, https://apps.dtic.mil/sti/trecms/pdf/AD1183861.pdf

37. Advancing Arctic sea ice remote sensing with AI and deep learning: now and future, accessed on May 18, 2025, https://egusphere.copernicus.org/preprints/2024/egusphere-2023-2831/

38. Maintain Optimal Performance in Robotics with Environmental Adaptation - Boston Engineering, accessed on May 18, 2025, https://blog.boston-engineering.com/maintain-optimal-performance-in-robotics-with-environmental-adaptation

39. A survey on autonomous environmental monitoring approaches: towards unifying active sensing and reinforcement learning - Frontiers, accessed on May 18, 2025, https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2024.1336612/full

40. Adaptive Robot Localization with Ultra-wideband Novelty Detection - arXiv, accessed on May 18, 2025, https://arxiv.org/html/2505.05903v1

41. Novelty Detection as an Intrinsic Motivation for Cumulative Learning Robots - CiteSeerX, accessed on May 18, 2025, https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=f4cdc976df21feeb3e465496818ac5e6a23584b5

42. Novelty detection in rover-based planetary surface images using autoencoders - Frontiers, accessed on May 18, 2025, https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2022.974397/full

43. ai.jpl.nasa.gov, accessed on May 18, 2025, https://ai.jpl.nasa.gov/public/documents/papers/wagstaff-ijcai2017-novelty.pdf

44. Novelty detection in rover-based planetary surface images using ..., accessed on May 18, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC9613947/

45. Open-World Novelty Detection, Characterization, and Accommodation - HRILab Tufts, accessed on May 18, 2025, https://hrilab.tufts.edu/projects/openworld/

46. Signal Novelty Detection as an Intrinsic Reward for Robotics - ResearchGate, accessed on May 18, 2025, https://www.researchgate.net/publication/370041160_Signal_Novelty_Detection_as_an_Intrinsic_Reward_for_Robotics

47. Self-Healing AI Systems: Autonomous Detection and Recovery from Security Breaches, accessed on May 18, 2025, https://www.researchgate.net/publication/390288670_Self-Healing_AI_Systems_Autonomous_Detection_and_Recovery_from_Security_Breaches

48. Self-Healing AI Systems: How Autonomous AI Agents Detect, accessed on May 18, 2025, https://aithority.com/machine-learning/self-healing-ai-systems-how-autonomous-ai-agents-detect-prevent-and-fix-operational-failures/

49. Self-Healing Cloud Systems: Integrating AI for Automated Recovery - ResearchGate, accessed on May 18, 2025, https://www.researchgate.net/publication/391424429_Self-Healing_Cloud_Systems_Integrating_AI_for_Automated_Recovery/download

50. Functional Self-Awareness and Metacontrol for Underwater Robot ..., accessed on May 18, 2025, https://www.mdpi.com/1424-8220/21/4/1210

51. Fast SEU Detection and Recovery in FPGA- Based AI Accelerators - Indico at ESA / ESTEC, accessed on May 18, 2025, https://indico.esa.int/event/531/contributions/10620/attachments/6614/11764/SEFUW_Eleonora_Vacca.pdf

52. STRAIT: Self-Test and Self-Recovery for AI Accelerator - ResearchGate, accessed on May 18, 2025, https://www.researchgate.net/publication/367138397_STRAIT_Self-Test_and_Self-Recovery_for_AI_Accelerator

53. Autonomous Agents in Energy Management - SmythOS, accessed on May 18, 2025, https://smythos.com/ai-industry-solutions/energy/autonomous-agents-in-energy-management/

54. Agentic AI in Energy Sector: Pioneering Autonomous Energy Intelligence - XenonStack, accessed on May 18, 2025, https://www.xenonstack.com/blog/agentic-ai-energy-sector

55. Revolutionizing AI Robotics: Ferri Solutions for Real-Time Data and Durability, accessed on May 18, 2025, https://www.siliconmotion.com/company/blog/Revolutionizing_AI_Robotics/detail

56. Trend 2025: Energy requirements often depend on the size of the AI ..., accessed on May 18, 2025, https://www.roboticstomorrow.com/story/2025/01/trend-2025-energy-requirements-often-depend-on-the-size-of-the-ai-model-/24021/

57. Onboard Optimization and Learning: A Survey - arXiv, accessed on May 18, 2025, https://arxiv.org/html/2505.08793v1

58. (PDF) Onboard Optimization and Learning: A Survey - ResearchGate, accessed on May 18, 2025, https://www.researchgate.net/publication/391741515_Onboard_Optimization_and_Learning_A_Survey

59. Industry Insights: How to Effectively Manage Robot Energy ..., accessed on May 18, 2025, https://www.automate.org/industry-insights/how-to-effectively-manage-robot-energy-consumption

60. In-Situ Resource Utilization (ISRU) - Meegle, accessed on May 18, 2025, https://www.meegle.com/en_us/topics/space-commercial/in-situ-resource-utilization-isru

61. In-situ Resource Utilisation Facility - CSIRO, accessed on May 18, 2025, https://www.csiro.au/en/work-with-us/use-our-labs-facilities/ISRU-Facility

62. Autonomous Systems NASA Capability Overview, accessed on May 18, 2025, https://www.nasa.gov/wp-content/uploads/2015/03/nac_tie_aug2018_tfong_tagged.pdf

63. ojs.aaai.org, accessed on May 18, 2025, https://ojs.aaai.org/index.php/AAAI/article/view/35208/37363

64. (PDF) Integrating AI Models for Voltage and Current Monitoring in ..., accessed on May 18, 2025, https://www.researchgate.net/publication/390722085_Integrating_AI_Models_for_Voltage_and_Current_Monitoring_in_Autonomous_Mobile_Robots_to_Prevent_Power_System_Blackouts/download

65. www.managementsolutions.com, accessed on May 18, 2025, https://www.managementsolutions.com/sites/default/files/minisite/static/22959b0f-b3da-47c8-9d5c-80ec3216552b/iax/pdf/explainable-artificial-intelligence-en.pdf

66. Trustworthy XAI and Application - arXiv, accessed on May 18, 2025, https://arxiv.org/html/2410.17139

67. ABOR Regents' Grant boosts AI-powered disaster response efforts ..., accessed on May 18, 2025, https://www.azregents.edu/news-releases/abor-regents-grant-boosts-ai-powered-disaster-response-efforts-help-keep-arizona

68. personalpages.manchester.ac.uk, accessed on May 18, 2025, https://personalpages.manchester.ac.uk/staff/louise.dennis/pubs/PID5539853.pdf

69. Empowering Safety-Critical Systems with Formal Methods., accessed on May 18, 2025, https://www.trust-in-soft.com/resources/blogs/how-formal-methods-improves-the-verification-of-safety-critical-systems

70. November 2023: How is AI being used for good in the Arctic ..., accessed on May 18, 2025, https://www.bas.ac.uk/media/beyondtheice/november-2023-how-is-ai-being-used-for-good-in-the-arctic/

71. Verification for space robotics - The University of Manchester, accessed on May 18, 2025, https://personalpages.manchester.ac.uk/staff/louise.dennis/pubs/iet_books.24287.011.pdf

72. Challenges in Autonomous Robotic System Verification - CEUR-WS.org, accessed on May 18, 2025, https://ceur-ws.org/Vol-3860/paper_7.pdf

73. Formal Verification of Real-Time Autonomous Robots: An Interdisciplinary Approach, accessed on May 18, 2025, https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2022.791757/full

74. Using Formal Methods for Autonomous Systems: Five Recipes for Formal Verification - arXiv, accessed on May 18, 2025, https://arxiv.org/pdf/2012.00856

75. A Runtime Safety Monitoring Approach for Adaptable Autonomous Systems - ResearchGate, accessed on May 18, 2025, https://www.researchgate.net/publication/335557336_A_Runtime_Safety_Monitoring_Approach_for_Adaptable_Autonomous_Systems

76. Runtime Assurance and Enforcement - Sebastian Junges, accessed on May 18, 2025, https://sjunges.github.io/research/runtime

77. A Case Study on Runtime Monitoring of an Autonomous Research Vehicle (ARV) System - Electrical and Computer Engineering, accessed on May 18, 2025, https://users.ece.cmu.edu/~koopman/pubs/kane15_monitoring.pdf

78. Identifying Run-time Monitoring Requirements for Autonomous Systems through the Analysis of Safety Arguments, accessed on May 18, 2025, https://eprints.whiterose.ac.uk/id/document/2846806

79. Leadership Ethics: Guide to Ethical Decision-Making - Hyperspace, accessed on May 18, 2025, https://hyperspace.mv/leadership-ethics/

80. Ethical strategy: How to align your business vision and mission with ethical principles and goals - FasterCapital, accessed on May 18, 2025, https://fastercapital.com/content/Ethical-strategy--How-to-align-your-business-vision-and-mission-with-ethical-principles-and-goals.html

81. Ethical Alignment In Leadership Decisions - Aurora Training Advantage, accessed on May 18, 2025, https://auroratrainingadvantage.com/leadership/ethical-alignment-leadership-decisions/

82. Misalignments in AI Perception: Quantitative Findings and Visual Mapping of How Experts and the Public Differ in Expectations and Risks, Benefits, and Value Judgments - arXiv, accessed on May 18, 2025, https://arxiv.org/html/2412.01459v1

83. AI value alignment: Aligning AI with human values | World Economic ..., accessed on May 18, 2025, https://www.weforum.org/stories/2024/10/ai-value-alignment-how-we-can-align-artificial-intelligence-with-human-values/

84. Supervisory Control as a Pathway to Smarter Robots | ETF Trends, accessed on May 18, 2025, https://www.etftrends.com/disruptive-technology-channel/supervisory-control-pathway-smarter-robots/

85. cognizancejournal.com, accessed on May 18, 2025, https://cognizancejournal.com/vol5issue3/V5I328.pdf

86. Optimizing Human-AI Collaboration: A Guide to HITL, HOTL, and HIC Systems, accessed on May 18, 2025, https://www.deepscribe.ai/resources/optimizing-human-ai-collaboration-a-guide-to-hitl-hotl-and-hic-systems

87. cdn.aaai.org, accessed on May 18, 2025, https://cdn.aaai.org/Symposia/Spring/1999/SS-99-06/SS99-06-013.pdf

88. mers-papers.csail.mit.edu, accessed on May 18, 2025, http://mers-papers.csail.mit.edu/Conference/2012/bushEtAl_IEEE2012/06187312.pdf

89. Experiments in Adjustable Autonomy, accessed on May 18, 2025, https://faculty.cs.byu.edu/~mike/mikeg/papers/IJCAI01.pdf

90. www.ri.cmu.edu, accessed on May 18, 2025, https://www.ri.cmu.edu/pub_files/pub4/steinfeld_aaron_m_2006_1/steinfeld_aaron_m_2006_1.pdf

91. Alternative Interfaces for Human-initiated Natural Language Communication and Robot-initiated Haptic Feedback: Towards Better Situational Awareness in Human-Robot Collaboration - arXiv, accessed on May 18, 2025, https://arxiv.org/html/2401.13903v1

92. (PDF) GENERATIVE AI FOR SCIENTIFIC DISCOVERY ..., accessed on May 18, 2025, https://www.researchgate.net/publication/390371479_GENERATIVE_AI_FOR_SCIENTIFIC_DISCOVERY_AUTOMATED_HYPOTHESIS_GENERATION_AND_TESTING

93. Accelerating scientific breakthroughs with an AI co-scientist, accessed on May 18, 2025, https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/

94. Agentic AI for Scientific Discovery, accessed on May 18, 2025, https://aaaiagenticai.github.io/

95. AstroAgents: A Multi-Agent AI for Hypothesis Generation from Mass ..., accessed on May 18, 2025, https://astrobiology.com/2025/04/astroagents-a-multi-agent-ai-for-hypothesis-generation-from-mass-spectrometry-data.html

96. AstroAgents: A Multi-Agent AI for Hypothesis Generation from Mass Spectrometry Data, accessed on May 18, 2025, https://www.researchgate.net/publication/390355098_AstroAgents_A_Multi-Agent_AI_for_Hypothesis_Generation_from_Mass_Spectrometry_Data

97. Literature Meets Data: A Synergistic Approach to Hypothesis Generation, accessed on May 18, 2025, https://chicagohai.github.io/hypogenic-demo/

98. Study explores the use of robots and artificial intelligence to ..., accessed on May 18, 2025, https://www.plymouth.ac.uk/news/study-explores-the-use-of-robots-and-artificial-intelligence-to-understand-the-deep-sea

99. Need a research hypothesis? Ask AI. | MIT News | Massachusetts Institute of Technology, accessed on May 18, 2025, https://news.mit.edu/2024/need-research-hypothesis-ask-ai-1219

100. AI in Space: Exploration, Research, Innovation, and Inclusivity | American Public University, accessed on May 18, 2025, https://www.apu.apus.edu/area-of-study/math-and-science/resources/ai-in-space/

101. The Future of AI Agents in Space Exploration - OKMG, accessed on May 18, 2025, https://www.okmg.com/blog/the-future-of-ai-agents-in-space-exploration

102. A Helping Hand: A Survey About AI-Driven Experimental Design for ..., accessed on May 18, 2025, https://www.mdpi.com/2076-3417/15/9/5208

103. AI-Powered “Self-Driving” Labs: Accelerating Life Science R&D ..., accessed on May 18, 2025, https://www.scispot.com/blog/ai-powered-self-driving-labs-accelerating-life-science-r-d

104. Autonomous Discovery | Argonne National Laboratory, accessed on May 18, 2025, https://www.anl.gov/autonomous-discovery

105. Science Use Case Design Patterns for Autonomous Experiments - OSTI.GOV, accessed on May 18, 2025, https://www.osti.gov/servlets/purl/2301625

106. The future of self-driving laboratories: from human in the loop interactive AI to gamification, accessed on May 18, 2025, https://pubs.rsc.org/en/content/articlehtml/2024/dd/d4dd00040d

107. LLM-Driven Autonomous Scientific Discovery - YouTube, accessed on May 18, 2025, https://www.youtube.com/watch?v=_rf3nI_U3Gk

108. Lessons Learned from Autonomous Sciencecraft Experiment - Artificial Intelligence Group, accessed on May 18, 2025, https://ai.jpl.nasa.gov/public/documents/papers/chien-aamas2005-lessons.pdf

109. New AI Model Is a Leap for Autonomous Materials Science | Feature ..., accessed on May 18, 2025, https://www.pnnl.gov/news-media/new-ai-model-leap-autonomous-materials-science

110. Fundamentals of AI in Scientific Research - NCBI, accessed on May 18, 2025, https://www.ncbi.nlm.nih.gov/books/NBK603480/

111. Advancing Human-Machine Teaming: Concepts, Challenges, and Applications - arXiv, accessed on May 18, 2025, https://arxiv.org/html/2503.16518v2

112. Advancing Human-Machine Teaming: Concepts, Challenges, and Applications - arXiv, accessed on May 18, 2025, https://arxiv.org/pdf/2503.16518

113. Human–Autonomy Teaming: A Review and Analysis of the Empirical Literature - PMC, accessed on May 18, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC9284085/

114. Calibrating Agency: Human-Autonomy Teaming and the Future of Work amid Highly Automated Systems - EPIC, accessed on May 18, 2025, https://www.epicpeople.org/calibrating-agency-human-autonomy-teaming-future-work-automated-systems/

115. arxiv.org, accessed on May 18, 2025, https://arxiv.org/pdf/1905.07379

116. core-robotics.gatech.edu, accessed on May 18, 2025, https://core-robotics.gatech.edu/files/2023/09/human_robot_teaming.pdf

117. A Framework for Dynamic Situational Awareness in Human Robot Teams: An Interview Study - arXiv, accessed on May 18, 2025, https://arxiv.org/html/2501.08507v1

118. ntrs.nasa.gov, accessed on May 18, 2025, https://ntrs.nasa.gov/api/citations/20190001937/downloads/20190001937.pdf

119. Supervised Autonomy for Exploration and Mobile Manipulation in Rough Terrain with a Centaur-Like Robot - Frontiers, accessed on May 18, 2025, https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2016.00057/full

120. (PDF) Adaptive trust calibration for human-AI collaboration, accessed on May 18, 2025, https://www.researchgate.net/publication/339424118_Adaptive_trust_calibration_for_human-AI_collaboration

121. A Quantum Model of Trust Calibration in Human–AI Interactions, accessed on May 18, 2025, https://www.mdpi.com/1099-4300/25/9/1362

122. Trust and performance in human-AI systems for multi-domain command and control - MIT Lincoln Laboratory, accessed on May 18, 2025, https://www.ll.mit.edu/sites/default/files/publication/doc/trust-performance-human-ai-systems-multi-domain-pentland-116777.pdf

123. Adaptive trust calibration for human-AI collaboration | PLOS One, accessed on May 18, 2025, https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0229132

124. PINEBERRY: Building Secure and Explainable AI for Space ..., accessed on May 18, 2025, https://www.kplabs.space/news/pineberry-building-secure-and-explainable-ai-for-space-missions

125. Explained: Generative AI's environmental impact | MIT News, accessed on May 18, 2025, https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117

126. AI has an environmental problem. Here's what the world can do about that. - UNEP, accessed on May 18, 2025, https://www.unep.org/news-and-stories/story/ai-has-environmental-problem-heres-what-world-can-do-about

127. AI IN UNDERWATER EXPLORATION: MAPPING THE OCEAN'S SECRETS INUA AI, accessed on May 18, 2025, https://inuaai.com/edge-ai-decentralizing-intelligence-for-faster-computing-copy/

128. AI in Space Operations: Opportunities and Challenges - Observer Research Foundation, accessed on May 18, 2025, https://www.orfonline.org/research/ai-in-space-operations-opportunities-and-challenges

129. Robotics and autonomy: From space robots to applications in daily life - ResearchPod, accessed on May 18, 2025, https://researchpod.org/informatics-technology/robotics-autonomy-space-robots-applications-daily-life

130. A Critical AI View on Autonomous Vehicle Navigation: The Growing Danger - MDPI, accessed on May 18, 2025, https://www.mdpi.com/2079-9292/13/18/3660

131. News: The role of AI and robotics in advancing space exploration and satellite deployment, accessed on May 18, 2025, https://www.automate.org/news/-68