Space robotics & sustainability: an agentic AI architecture for autonomous debris removal as case study

Space robotics & sustainability: an agentic AI architecture for autonomous debris removal as case study Maldonado-Romo, Javier Hyoksi Tech [email protected] Abstract Communication latency in space operations limits continuous human supervision and real-time decision-making. This paper presents a delay-tolerant agentic AI architecture that enables autonomous, sustainable, and explainable decision processes for space robotics. The framework integrates perception, reasoning, governance, and sustainability agents operating directly onboard the spacecraft. During communication gaps, the system maintains local autonomy and logs interpretable reasoning chains for later human validation. Once connectivity is restored, operators can review and refine these decisions through a virtual reality interface, enabling asynchronous supervision. The architecture mitigates latency-related risks by embedding ethical reasoning and sustainability assessment within the onboard decision flow. A Sustainability Impact Score (SIS) guides adaptive agent prioritization by balancing energy efficiency, mission safety, and environmental impact. Simulations in virtual environments demonstrate gains in mission resilience, transparency, and sustainability performance. The paper introduces Space Robotics & Sustainability as an emerging paradigm that unifies autonomous operations and sustainability principles into a coherent computational framework for 1 decision-making. Keywords: Human-in-the-Loop Systems; Virtual Simulation; Autonomous DecisionMaking ; Sustainable Space Mission 1 Introduction The growing congestion of Earth’s orbital environment has made the long-term sustainability of space operations increasingly uncertain. Decades of satellite deployment, accidental breakups, and uncontrolled fragmentation events have filled near-Earth space with tens of thousands of traceable objects and countless smaller fragments. As the density of this material rises, so does the likelihood that a single collision could initiate chain reactions that endanger operational satellites, human missions, and future infrastructure, a phenomenon widely associated with the Kessler Syndrome (Kessler & Cour-Palais, 1978). Ensuring safe access to space now depends on strategies capable of actively reducing the debris population rather than merely containing its growth. Among these strategies, Active Debris Removal (ADR) has become a focal point of international attention. Yet, executing ADR missions remains non-trivial: non-cooperative targets often rotate unpredictably, travel at high relative velocities, and occupy diverse orbital regimes that challenge perception, guidance, and control systems (Shan et al., 2016). Compounding the technical difficulty, ADR operations must satisfy an expanding set of regulatory expectations from ISO 24113 debris-mitigation requirements (ISO, 2023) to national licensing provisions and UN sustainability guidelines (UNOOSA, 2021). These constraints demand mission designs that are technically sound and fully accountable within evolving legal and ethical frameworks. However, current ADR architectures, tend to rely on deterministic scripts or tightly predefined behaviors that struggle to adapt when real-world conditions deviate from assumptions. Limited onboard intelligence and weak integration of sustainability considerations often result in rigid systems unable to cope with dynamic debris motion or unexpected 2 sensor degradation (Jaekel et al., 2018). Furthermore, high operational costs and the absence of economically viable service models continue to hinder widespread adoption (Liou & Johnson, 2009). As space traffic intensifies and disposal regulations tighten, for example, the U.S. FCC’s five-year deorbit rule for Low Earth Orbit (LEO) spacecraft and ESA’s growing emphasis on debris-removal services (Bennett, 2025). Therefore, the need for autonomous, regulation-aligned mission architectures becomes increasingly urgent (Y.-H. Jiang et al., 2025). Recent advances in autonomous systems and multi-agent AI provide a promising foundation for addressing these challenges. Agentic AI architectures composed of specialized modules capable of sensing, planning, coordination, communication, and ethical reasoning (Picard et al., 2021) enable flexible, scalable autonomy that can embed governance and sustainability into the core of the decision-making process. These developments create opportunities for ADR missions to operate more adaptively and transparently, even in uncertain or communication-limited scenarios. Building on this technological trajectory, the present document introduces a system that integrates agentic AI modules capable of maintaining situational awareness, logging interpretable decision traces, and evaluating sustainability impacts as part of routine operations. Once ground contact resumes, a virtual reality off line interface allows operators to review decisions asynchronously, refine policies, and ensure compliance without the need for continuous supervision. In contrast to static or rule-based ADR approaches, the proposed framework incorporates a VR-driven closed loop training mechanism in which large language model (LLM) reasoning agents learn from diverse, high-fidelity simulations to generate context aware strategies that generalize beyond nominal conditions. This hybrid AI-human workflow enhances robustness, reduces operational risk, and strengthens the transparency of autonomous actions. The framework is organized into a set of specialized layers ranging from perception and prediction to planning, governance, ethics, learning, and simulation each populated by agents 3 with targeted responsibilities. A Sustainability Impact Score (SIS) guides the prioritization of actions by balancing fuel usage, risk mitigation, regulatory compliance, environmental impact, and the degree of human intervention. This combination of hierarchical autonomy, sustainability aware evaluation, and VR-enabled learning forms the basis of a new conceptual paradigm that is refer as Space Robotics & Sustainability. To demonstrate the applicability of the framework, three representative ADR missions, such as Astroscale’s ELSA-d (Astroscale, 2018), ESA’s ClearSpace-1 (ClearSpace, 2019; Vasconcelos et al., 2025), and the Surrey Space Centre’s RemoveDEBRIS experiment (Forshaw et al., 2016, 2017, 2020; Aglietti et al., 2020) are examined as operational case studies. These missions provide varied scenarios involving magnetic docking, robotic capture, and net or harpoon based techniques, offering distinct opportunities to validate the SIS based prioritization and evaluate the strengths of agentic reasoning under different operational constraints. These missions serve as realistic testbeds for simulation-driven evaluation and VR-based training. The structure of the paper is as follows. Section 2 presents the relevant background. Section 3 describes the agentic AI architecture, its layers, and the SIS. Section 4 examines the integration of the framework with the selected ADR missions. Section 5 discusses implications, challenges, and opportunities, and Section 6 concludes with remarks on sustainable, autonomous mission design. 2 Background The landscape of autonomous space robotics has shifted from single-purpose, manually scripted systems to increasingly interconnected robotic teams capable of executing cooperative tasks in orbit and on planetary surfaces. Demonstrations such as ESA’s ARCHES initiative and JAXA’s MMX mission illustrate this transition, showcasing combinations of landers, rovers, and manipulators operating jointly in complex environments through stereo 4 imaging, cooperative mapping, sample acquisition, and multi-sensor coordination techniques (Forshaw et al., 2016, 2017; Stolfi & Danoy, 2024). These platforms rely on advanced localization pipelines ranging from stereo vision and 6D SLAM to carrier-based navigation and high-resolution sensing jkfor accurate scene interpretation and target characterization (Capra et al., 2023). Nevertheless, despite these capabilities, their decision processes typically follow predefined routines and lack adaptable prioritization mechanisms or sustainability-aware logic. Research in agent-based systems has sought to address such limitations by introducing distributed reasoning, modularity, and reconfigurable intelligence into cyber-physical architectures. Model-driven and ontology-centered engineering approaches (Capra et al., 2023; Xu et al., 2023) allow heterogeneous robotic elements to coordinate more flexibly, while formal meta-models (Haddad et al., 2021; Sumiea et al., 2023) have been used to establish structured rules for agent behavior, sequencing, and communication. In mission planning domains, multi-agent strategies have informed hierarchical task allocation and adaptive scheduling for lunar construction and assembly missions, enabling systems to manage uncertainty and operator preferences (Ribeiro & Gomes, 2021; Wang & Chen, 2025). However, these frameworks rarely incorporate sustainability-driven criteria or detailed human oversight when applied to on-orbit servicing or debris-removal operations. Parallel advances in LLMs and multimodal variants have introduced new opportunities for natural language interaction, spatially grounded reasoning, and high-level perception in robotic systems (Christensen et al., 2025; Ghiani et al., 2024). Vision language models (VLMs) capable of jointly interpreting imagery, text, and spatial context (Frering et al., 2025; Killeen et al., 2025) offer a foundation for explainable autonomy during safety tasks. Recent studies demonstrate how autonomous agents can translate human instructions into executable actions (Qi et al., 2026; Zhang et al., 2025) and share information cooperatively. Yet, the potential of such models for space debris removal where operations require reliability under uncertainty and strict safety constraints remains insufficiently explored. 5 Sustainability has also become a prominent design requirement for orbital missions, driven by the rapid growth of mega-constellations and the global push for long-term stewardship of near Earth space. Frameworks adapted from the triple bottom line encompassing environmental, social, and economic considerations have been extended to evaluate space activities (McKnight et al., 2021; Nosseir et al., 2021). Regulatory instruments such as the UN COPUOS guidelines promote responsible operational practices and emphasize sustainability throughout mission lifecycles (Pardini & Anselmo, 2021; Murtaza et al., 2020). However, practical challenges persist, including reconciling operational efficiency with stewardship obligations, addressing the "space sustainability paradox" (Migaud, 2020), and embedding circular economy principles into spacecraft design, reuse, and end-of-life strategies (Wilson & Vasile, 2023; Stokes et al., 2020). High-fidelity virtual environments and simulation frameworks have become crucial to evaluating autonomous behaviors under realistic orbital conditions. These platforms enable controlled replication of orbital mechanics, illumination regimes, sensor performance, and debris dynamics that are costly or impossible to reproduce physically (Maldonado-Romo et al., 2022). Consequently, VR-based testing supports safe evaluation of capture strategies, multi-agent coordination, and the interplay between risk, fuel expenditure, and sustainability considerations. It also facilitates rapid prototyping and stress testing of decision-making algorithms in uncertain or degraded scenarios. When integrated with LLM-powered reasoning and sustainability aware agent architectures, VR becomes a powerful iterative development tool. It enables simulated human-robot supervision, provides opportunities to refine agent policies, and allows dynamic tuning of metrics such as the Sustainability Impact Score (SIS). Despite its demonstrated effectiveness in Earth based robotics and planetary exploration, the combined use of VR, LLM-driven reasoning, and sustainability oriented multi-agent control in the context of active debris removal remains largely underdeveloped. 6 3 Methodology Space Robotics & Sustainability (SRS) presents a unified perspective in which autonomy, AI, and environmental responsibility are treated as interconnected pillars of orbital mission design. Whereas conventional space robotics traditionally centers on mechanical control, guidance strategies, and independent navigation (Nesnas et al., 2021), the SRS approach extends this foundation by incorporating governance, ethical safeguards, and sustainability criteria directly into the cognitive structure of the system. Similarly, sustainability-focused robotics research often emphasizes materials selection, end-of-life planning, or propulsion efficiency (Wilson & Vasile, 2023), but SRS reframes sustainability as an intrinsic decisionmaking principle, one that influences how agents interpret information, learn from experience, communicate with operators, and select actions. Under this paradigm, autonomy becomes a form of responsible agency in which onboard decisions are continuously aligned with safety requirements, mission rules, and long-term orbital stewardship. The agentic AI architecture proposed in this work brings layered autonomy, adaptive learning processes, policy and sustainability aware decision filters, and simulation-driven refinement to meet the challenges of active debris removal. The framework integrates four foundational elements: (1) a hierarchical multi-agent design inspired by classical robotic control stacks, (2) a prioritization mechanism informed by sustainability metrics, (3) human-aligned reasoning enabled through LLM interfaces, and (4) high-fidelity simulation environments in VR that support iterative training and validation. The subsections that follow elaborate on each of these components and describe how they operate collectively to support robust, transparent, and sustainable mission execution. 3.1 Agentic AI Framework The proposed agentic AI framework functions as a hierarchical and recursive architecture in which data, decisions, constraints, and learning signals continuously flow across layers. This 7 design enables safe, adaptive, and regulation-compliant ADR missions. 3.1.1 Layer: Sensing & Perception The Sensing & Perception layer serves as the entry point for information about the orbital environment, delivering the raw measurements and contextual cues needed for safe navigation and informed decision-making. Agents at this level are tasked with detecting, localizing, and characterizing debris in a dynamic, cluttered region of space. A diverse suite of sensing modalities can be deployed here, including optical cameras, radar, LiDAR, and infrared instruments. Each technology offers distinct benefits depending on target properties and environmental conditions. For instance, (Priyadarshini, 2024) shows that optical systems are well suited to tracking centimeter scale debris under favorable illumination, whereas radar-based approaches are better matched to long-range, all-weather observation and surveillance (Hao et al., 2021). To reduce uncertainty and improve robustness, perception agents employ data fusion mechanisms that combine complementary sensor streams. Kalman filter-based fusion strategies, such as those described in (S. Jiang et al., 2025), blend visual and radar information to maintain continuous tracking through occlusions or temporary sensor dropouts. As highlighted in (Lagona et al., 2022), aligning and time-synchronizing heterogeneous sensors is particularly challenging when dealing with rapidly moving debris. Operational constraints further complicate this layer. Perception agents must function under limited onboard computing budgets and in noisy conditions introduced by vibration, radiation, and long-term sensor degradation (Rech, 2024). Redundancy and fault tolerance are essential to mitigate failures, especially in LEO, where sensor bleaching or temporary blinding can occur during solar events (ERTEKIN et al., 2024). More advanced configurations incorporate lightweight AI-based feature extraction methods. As noted in (Perruci & Lee, 2025; Chai et al., 2021), convolutional filters and compact neural networks can run directly onboard to perform tasks such as edge detection, shape 8 inference, and pose estimation, thereby reducing the need to transmit raw data and lowering bandwidth demands. Agents in this layer focus on acquiring and preprocessing sensor streams to derive debris pose, velocity, and classification attributes. Sensor Fusion and Pose Estimation Agents, in particular, are key to maintaining reliable situational awareness under high relative velocities and degraded visibility. The resulting processed outputs provide the first semantic representation of the environment for the rest of the architecture. 3.1.2 Layer: Interpretation & Prediction Together, the Interpretation & Prediction layers constitute the cognitive core of the agentic AI framework. Their primary role is to transform processed perceptual inputs into missionrelevant assessments and strategies. In debris dense scenarios, the agents operating in these layers must work in close coordination. This layer focuses on turning sensor observations into a structured understanding of debris behavior. Agents estimate object pose, forecast trajectories, classify debris types, and assess risks such as uncontrolled tumbling, fragmentation, or potential collisions. Techniques described in (Hussain et al., 2025) combine Kalman filtering with learned motion models to anticipate the behavior of non-cooperative targets exhibiting complex spin or drift dynamics. Hybrid approaches that integrate classification networks with optical flow tracking (Nguyen et al., 2024) are used to improve prediction performance when observations are incomplete or noisy. Classification agents identify and categorize debris by form and function. As reported in (Liu et al., 2025), shape-based recognition models trained on synthetic imagery can distinguish between structural panels, tanks, and composite fragments. In (Caon et al., 2023), learned behavioral models are used to infer how debris may respond to interactions, facilitating assessments of stability and maneuverability post-contact. Once the state and behavior of objects have been inferred, navigation-focused agents con9 loops (May et al., 2022) to adjust decision logic dynamically as mission experience accumulates. 3.1.8 Layer: Ethics & Constraints The Ethics & Constraints layer encodes high-level ethical principles, safety rules, and operational boundaries that govern autonomous behavior. Its purpose is to ensure that agent decisions remain within acceptable risk and responsibility limits during debris removal operations. According to (Barra et al., 2025), agentic AI systems in space must adhere to strict constraints to avoid hazardous or undesirable outcomes, such as intentional or accidental collisions, incursion into protected orbital zones, or interference with sovereign spacecraft. Agents in this layer implement fail-safe logic and monitor conditions under which operations should be halted. They also interpret legal and ethical prescriptions in terms of concrete operational constraints, limiting actions toward non-cooperative or sensitive objects and enforcing data-handling and interoperability requirements in multinational missions. These constraints operate in parallel with technical planning and control layers, continually checking for deviations from defined safety margins. If a proposed action risks violating a rule or threshold, it is either modified or blocked in real time (Y.-H. Jiang et al., 2025). Typical agents include Fail-safe Enforcement Agents and Justification Agents, which scrutinize decisions for safety compliance and generate explanations when interventions occur. They help maintain alignment between agent behavior, human values, and international space norms. 3.1.9 Layer: Simulation & Evaluation The Simulation & Evaluation layer underpins the virtual testing and validation of the entire agentic AI stack in high-fidelity environments (Maldonado-Romo & Aldape-Pérez, 2021). Before deployment, agents across all layers perception, planning, actuation, communication, 16 sustainability, and ethics are exercised in realistic, closed-loop simulations to verify safety, compatibility, and performance. Simulation platforms recreate critical aspects of the operational context, including microgravity dynamics, debris motion, orbital parameters, and inter-agent interactions. They enable reinforcement-learning agents to train in photorealistic scenes with realistic sensor noise, latency, and control imperfections (Miyoshi et al., 2025). Scenario Evaluation Agents and Model Validator Agents coordinate the design and execution of these virtual campaigns. They synthesize representative debris fields, assess the behavior of multiple agents working together, and quantify robustness against off-nominal events. Synthetic failure cases generated in VR can be used to fine-tune LLM-based reasoning agents, improving sample efficiency and overall reliability. 3.2 VR simulation and human-in-the-loop integration Across all layers, agents record decisions, inputs, and outcomes, creating detailed logs of their internal reasoning processes. LLM components ingest these logs and summarize them into natural-language explanations that operators can inspect, enabling them to confirm agent behavior, diagnose anomalies, and revise policies. Human operators retain the ability to override or endorse autonomous choices, reinforcing accountability and trust. Virtual simulator reproduces sensor physics, orbital motion, lighting conditions, and communication delays, supporting scenarios ranging from nominal debris capture to operations in contested orbital regions or under degraded hardware conditions. Within this VR context, multiple tasks can be carried out: generating decision logs for LLM fine-tuning, training LLM agents using scenario-specific prompt templates, and reducing inference latency through preconditioned embeddings and structured context. Simulated human interaction sequences are also used to model realistic supervision patterns. These synthetic environments allow agents to be challenged under a broad variety of operational and environmental conditions, improving robustness and supporting continuous 17 retraining. The VR-LLM loop comprises four main stages: VR environment generation, in which debris distributions, lighting regimes, and policy constraints are instantiated; LLM fine-tuning, where simulated data and dialogue transcripts are used to specialize model behavior; Preconditioned Inference, which leverages scenario-specific embeddings and prompts for more efficient contextual reasoning; and real-world mission execution, where the refined models guide decisions with reduced computational overhead. Feedback from operations informs subsequent simulation cycles whenever performance gaps or new requirements are identified. This process enables the construction of domain specific embeddings and decision templates, lowers token-processing demands during in-mission human queries, and produces prevalidated response trees for high-risk decision branches. Through iterative VR-based training, reasoning agents gain transparency and operational effectiveness, supporting explainable autonomy even under strict resource constraints. To evaluate the effectiveness and sustainability of the agentic AI system within these VR experiments, a set of metrics is selected that reflects the objectives of real debris-removal missions and the principles of sustainable space operations, drawing on prior literature, mission reports, and established guidelines. 4 Case studies: mission-oriented agent prioritization and sustainable assessment To examine how the proposed agentic AI framework behaves under realistic conditions, this section considers three representative ADR missions: ELSA-d, ClearSpace-1, and RemoveDEBRIS. 18 4.1 Agentic AI configuration per mission For each mission scenario, the agent framework is tailored to the mission’s unique constraints and objectives. The specific set of agents and their relative importance depends on factors such as mass and power budgets, debris characteristics, autonomy requirements, and risk posture. This approach evaluates agents against sustainability-oriented and mission-specific criteria, and orders them according to their proximity to an ideal configuration. In practical terms, agents that strongly support mission goals, such as reliable capture, high energy efficiency, or effective risk mitigation are selected and emphasized, whereas agents with marginal impact are suppressed to conserve onboard resources. The three ADR missions complement roles as case studies. In ELSA-d, emphasis is placed on sensor fusion and attention mechanisms that support close-proximity guidance and magnetic docking. ClearSpace-1, by contrast, places higher priority on robust object recognition and sensing chains to guarantee dependable operation of its robotic capture system. RemoveDEBRIS highlights flexible, high-efficiency agents such as feature extraction and object recognition to support a portfolio of capture approaches (e.g., nets and harpoons). The resulting TOPSIS scores reflect how well each agent configuration approximates the “ideal” profile for a given mission. 4.2 VR simulation and human-in-the-loop integration Virtual scenarios are constructed to emulate the operational contexts of each mission and to evaluate both agent behavior and LLM-based reasoning interfaces. In the ELSA-d setting, simulations center on autonomous rendezvous and magnetic docking; for ClearSpace-1, they focus on multi-arm robotic capture and motion planning; and in the RemoveDEBRIS case, the emphasis is on net deployment and control under navigational uncertainty. Each mission scenario uses a customized weighting of objectives: ELSA-d prioritizes capture reliability, energy use, and risk reduction; ClearSpace-1 stresses compliance with constraints and component reuse; and RemoveDEBRIS highlights low-latency decision-making and collision risk 19 management. Within these VR environments, a continuous loop links agent decisions, SIS evaluation, and LLM training. Simulated outcomes are used to refine SIS values, capturing trade-offs among fuel consumption, safety margins, and environmental impact. At the same time, LLM reasoning agents are trained on mission-specific interaction patterns, which reduces model complexity and improves the efficiency of in-mission inference. Human-agent collaboration can be studied by imposing time pressure and ambiguity, allowing the evaluation of how effectively operators and agents coordinate under realistic constraints. This iterative VR-framework loop illustrates the benefits of simulation-led development for agentic AI in space operations. SIS can be recalibrated based on virtual mission statistics, enabling more nuanced sustainability-sensitive decision policies. Training on domain-specific data also constrains the LLM’s operating space, improving robustness and reducing token usage during live operations. In addition, virtual scenarios provide a safe testbed for analyzing how different supervision strategies ranging from passive monitoring to active intervention affect mission performance and explainability. In the ELSA-d simulation, the environment reproduces the end-of-life services demonstration conducted by Astroscale. A high-fidelity model of the chaser spacecraft includes a magnetic docking plate and articulated capture mechanisms, while the client satellite is rendered with surface degradation and reflective docking features to emulate realistic optical responses. The surrounding scene incorporates a detailed Earth model with dynamic clouds, varying illumination, and orbital eclipses. Additional small debris particles are injected to challenge perception agents. Synthetic LiDAR, optical imaging, and radar streams feed the perception pipeline, and relative dynamics follow Clohessy-Wiltshire motion to ensure physically consistent rendezvous behavior. Heads-up displays (HUDs) show ∆V, separation distance, and angular rates for human-in-the-loop monitoring. The ClearSpace-1 VR setup recreates the capture and disposal of the Vega Secondary Payload Adapter (Vespa). The Vespa object is modeled with accurate geometry and realistic 20 surface properties to support vision-based detection and tracking. The servicer spacecraft includes four articulated robotic arms with inverse kinematics for realistic grasp planning and execution. Target tumbling states can be varied to test the robustness of control agents under different dynamical conditions. Capture envelopes, contact forces, and approach trajectories are visualized to facilitate analysis. Fragment clouds are generated in the event of failed grasps to stress-test avoidance behaviors. The operator interface offers controls for LLM-based reasoning agents, enabling mission personnel to adjust, refine, or override agent decisions. For RemoveDEBRIS, the VR environment is designed around net and harpoon demonstrations. Debris objects are modeled with CubeSat-like dimensions and diverse geometries and reflectivities to challenge detection and tracking systems. The net deployment mechanism includes rope dynamics, collision handling, and entanglement behavior, allowing detailed assessment of capture stability. Optical and infrared sensing modes are supported to explore performance across lighting conditions. The visual background includes a realistic star field, atmospheric limb glow, and solar eclipse transitions, approximating the variability seen in orbit. 5 Discussion The findings of this work indicate that the proposed agentic AI framework can effectively manage the multifaceted demands of debris removal missions through configurations that are both autonomous and sustainability-aware. The framework offers a modular architecture that can be adapted to diverse mission profiles and operational environments by decomposing complex mission functionality into layered, specialized agents for sensing, planning, communication, governance, sustainability, ethics, and learning. The SIS cost function anchors agent selection in both performance and sustainability dimensions, including energy efficiency, risk reduction, component reusability, and cognitive 21 workload. This dual focus supports the alignment of operational decisions with long-term stewardship of the orbital environment and promotes sustainability as a core design driver rather than an afterthought. Results from VR-based experiments show that the tailored agentic configurations (A) consistently outperform baseline setups (B) across the studied missions. Simulated outcomes include reductions in fuel consumption on the order of 40-50%, shortened mission completion times of roughly 30%, and a decrease in human override interventions exceeding 60%. Such gains suggest that embedding sustainability-aware reasoning and delay-tolerant autonomy within the agent stack delivers tangible operational benefits. The SIS-based evaluation further ensures that decisions remain transparent and explainable while still improving safety and efficiency. The comparative analysis across ELSA-d, ClearSpace-1, and RemoveDEBRIS highlights recurring patterns. Some agents such as those focused on object recognition consistently emerge as high priority due to their foundational role in tracking and identification. Others, like Magnetic Docking Control agents, show high value only in missions that employ specific capture strategies. These patterns underscore the importance of configuring agent hierarchies according to mission objectives, available resources, and sustainability trade-offs rather than relying on a one-size-fits-all configuration. Mapping SIS contributions to individual agents creates a clear linkage between functional roles and sustainability outcomes. Agents that score highly across multiple SIS dimensions exert a larger systemic influence, simplifying decision-making around which capabilities to retain, enhance, or exclude in future designs. This mapping can inform mission planning, technology investment, and policy development for debris removal and orbital servicing. The framework’s modular design also facilitates scalability across different classes of missions. As new technologies, legal standards, and operational practices emerge, agents can be added, removed, or reconfigured without redesigning the entire architecture. This flexibility supports future multi-target debris campaigns, on-orbit servicing constellations, 22 and collaborative missions guided by sustainability metrics. Simultaneously, the human-inthe-loop functionality enabled by LLM interfaces introduces a semantic layer that allows mission planners to interrogate agent decisions, adjust constraints, and explore alternative strategies in natural language. This capability fosters transparency, strengthens ethical accountability, and builds confidence in autonomous systems. Thus, these elements illustrate how the SRS paradigm can serve as a structured basis for responsible and adaptive decision-making in communication-constrained settings. SRS reframes autonomy as a proactive governance mechanism that evaluates each action in ethical, technical, and environmental terms by merging agentic intelligence with sustainabilityoriented reasoning. This integration enables spacecraft to preserve operational integrity, traceability, and resource efficiency even when ground contact is intermittent or delayed. Moreover, it shifts sustainable decision-making from a post-mission evaluation metric to an operational advantage, outlining a replicable model for missions requiring distributed intelligence, extended autonomy, and asynchronous human validation. 6 Conclusion This work contributes to autonomous space robotics by introducing an agentic AI framework that embeds sustainability as a cognitive and operational principle. The proposed architecture weaves together reasoning, governance, learning, sustainability analysis, and simulation within a unified structure designed to support transparent, explainable, and regulation-aware decision-making. Through the Sustainability Impact Score (SIS), the framework promotes operations that are efficient, safe, and environmentally responsible, while remaining adaptable to a broad range of mission scenarios. Applying the framework to three emblematic ADR missions ELSA-d, ClearSpace-1, and RemoveDEBRIS demonstrates its ability to enhance fuel efficiency, reduce mission duration, and improve human-AI collaboration. The combination of VR-based simulation and 23 LLM-driven reasoning, grounded in physical constraints, provides a realistic and interactive environment for pre-mission evaluation. Within this environment, agents can develop robust, context-sensitive policies before being deployed in orbit. This research positions Space Robotics & Sustainability (SRS) as a practical and conceptual paradigm for designing intelligent missions. SRS enables spacecraft to operate responsibly under communication delays and incomplete information by integrating autonomy, ethics, and sustainability into a coherent decision framework. It offers a scalable template for missions involving distributed intelligence, cooperative servicing, and long-duration operations, where sustainability becomes a core driver of reasoning rather than an external constraint. Future work will extend this approach to constellations of services and multi-satellite coordination, with a focus on adaptive policy learning and real-time governance mechanisms for interplanetary and cislunar missions. These directions will further consolidate SRS as a reference framework for sustainable, autonomous, and ethically aligned space systems capable of making informed decisions under remote, latency-affected conditions. References Aglietti, G. S., Taylor, B., Fellowes, S., Salmon, T., Retat, I., Hall, A., ... Steyn, W. H. (2020, March). The active space debris removal mission removedebris. part 2: In orbit operations. Acta Astronautica,168, 310-322. doi: 10.1016/j.actaastro.2019.09.001 Astroscale. (2018). Elsa-d: An in-orbit end-of-life demonstration mission. https://astroscale.com/wp-content/uploads/2018/09/ELSA-1-Conference-IAC -2018-v1.1.pdf. (Accessed: 7 August 2025) Barea, A., Gonzalo, J. L., Colombo, C., & Urrutxua, H. (2024, October). A constraint programming framework for preliminary mission analysis: Applications for constellation24 servicing active debris removal. Advances in Space Research,74(7), 3060-3080. doi: 10.1016/j.asr.2024.07.015 Barra, F. L., Rodella, G., Costa, A., Scalogna, A., Carenzo, L., Monzani, A., & Corte, F. D. (2025, May). From prompt to platform: an agentic ai workflow for healthcare simulation scenario design. Advances in Simulation,10(1). doi: 10.1186/s41077-025-00357-z Bassetto, M., Mengali, G., & Quarta, A. A. (2024, December). Drag sail attitude tracking via nonlinear control. Acta Astronautica,225, 845-856. doi: 10.1016/j.actaastro.2024.09.046 Bennett, M. M. (2025, May). Orbital debris requires prevention and mitigation across the satellite life cycle. Communications Engineering,4(1). doi: 10.1038/s44172-025-00430-5 Blair, C. R. (1959, July). On computer transcription of manual morse. Journal of the ACM , 6(3), 429-442. doi: 10.1145/320986.320997 Bonnal, C., Francillout, L., Moury, M., Aniakou, U., Dolado Perez, J.-C., Mariez, J., & Michel, S. (2020, February). Cnes technical considerations on space traffic management. Acta Astronautica,167 , 296-301. doi: 10.1016/j.actaastro.2019.11.023 Caon, A., Branz, F., & Francesconi, A. (2023, September). Smart capture tool for space robots. Acta Astronautica,210, 71-81. doi: 10.1016/j.actaastro.2023.05.014 Capra, L., Brandonisio, A., & Lavagna, M. (2023, May). Network architecture and action space analysis for deep reinforcement learning towards spacecraft autonomous guidance. Advances in Space Research,71 (9), 3787-3802. doi: 10.1016/j.asr.2022.11.048 Cassinis, L. P., Park, T. H., Stacey, N., D’Amico, S., Menicucci, A., Gill, E., . .. SanchezGestido, M. (2023, June). Leveraging neural network uncertainty in adaptive unscented kalman filter for spacecraft pose estimation. Advances in Space Research,71(12), 50615082. Retrieved from http://dx.doi.org/10.1016/j.asr.2023.02.021 25 Pratt, M., Ameel, T., & Rao, S. R. (2025, May). Modeling of thermal enhancement and scaling analysis for omnidirectional magnetic field generator to actively detumble space debris. International Journal of Heat and Mass Transfer,241, 126733. doi: 10.1016/ j.ijheatmasstransfer.2025.126733 Priyadarshini, I. (2024, December). Enhanced space debris detection and monitoring using a hybrid bi-lstm-cnn and bayesian optimization. Artificial Intelligence and Applications, 3(1), 43-55. doi: 10.47852/bonviewaia42023741 Qi, J., Lu, L., Wang, F., Lee, H.-Y., Navarro-Alarcon, D., Zhang, Z., & Zhou, P. (2026, February). Llm-driven symbolic planning and hierarchical imitation learning for longhorizon deformable object assembly. Robotics and Computer-Integrated Manufacturing, 97, 103096. doi: 10.1016/j.rcim.2025.103096 Rafalskyi, D., Martínez, J. M., Habl, L., Zorzoli Rossi, E., Proynov, P., Boré, A., . .. Aanesland, A. (2021, November). In-orbit demonstration of an iodine electric propulsion system. Nature,599 (7885), 411-415. doi: 10.1038/s41586-021-04015-y Rai, M. C., Nair, M. H., Schaefer, D., Detry, R., Poozhiyil, M., Rybicka, J., . .. Gancet, J. (2024, November). Robotic upcycling and recycling: unraveling the era of sustainable in-space manufacturing. CEAS Space Journal,17(3), 455-469. doi: 10.1007/s12567-024 -00576-6 Rech, P. (2024, April). Artificial neural networks for space and safety-critical applications: Reliability issues and potential solutions. IEEE Transactions on Nuclear Science,71(4), 377-404. doi: 10.1109/tns.2024.3349956 Ribeiro, L., & Gomes, L. (2021). Describing structure and complex interactions in multiagent-based industrial cyber-physical systems. IEEE Access,9, 153126-153141. doi: 10 .1109/access.2021.3127344 32 Rovetto, R. J., Kelso, T., & O’Neil, D. A. (2020, September). Orbital debris ontology, terminology, and knowledge modeling. Journal of Space Safety Engineering,7(3), 451458. doi: 10.1016/j.jsse.2020.07.008 Sarego, G., Olivieri, L., Valmorbida, A., Brunello, A., Lorenzini, E. C., Tarabini Castellani, L., .. . Sánchez-Arriaga, G. (2021, February). Deployment requirements for deorbiting electrodynamic tether technology. CEAS Space Journal,13 (4), 567-581. doi: 10.1007/ s12567-021-00349-5 Sarritzu, A., & Pasini, A. (2024, April). Performance comparison of green propulsion systems for future orbital transfer vehicles. Acta Astronautica,217, 100-115. doi: 10.1016/ j.actaastro.2024.01.032 Shan, M., Guo, J., & Gill, E. (2016, January). Review and comparison of active space debris capturing and removal methods. Progress in Aerospace Sciences,80 , 18-32. doi: 10.1016/j.paerosci.2015.11.001 Simha, N., Servadio, S., Lifson, M., Lavezzi, G., & Linares, R. (2025, March). Optimal active debris removal mission planning to inform policy decisions. Acta Astronautica, 228, 224-236. doi: 10.1016/j.actaastro.2024.11.050 Skinner, M. A., Bates, B., Leonard, S., Neff, J., Bauer, P., von Tobel, B., . .. Feuge-Miller, B. (2023, September). Maneuvering into the future: Open-architecture data repository (oadr) prototype: Towards civil and commercial space traffic coordination. Journal of Space Safety Engineering,10(3), 366-373. doi: 10.1016/j.jsse.2023.07.002 Stokes, H., Akahoshi, Y., Bonnal, C., Destefanis, R., Gu, Y., Kato, A., .. . Tang, M. (2020, September). Evolution of iso’s space debris mitigation standards. Journal of Space Safety Engineering,7(3), 325-331. doi: 10.1016/j.jsse.2020.07.004 33 Stolfi, D. H., & Danoy, G. (2024, February). Evolutionary swarm formation: From simulations to real world robots. Engineering Applications of Artificial Intelligence,128, 107501. doi: 10.1016/j.engappai.2023.107501 Sumiea, E. H. H., Abdulkadir, S. J., Ragab, M. G., Al-Selwi, S. M., Fati, S. M., AlQushaibi, A., & Alhussian, H. (2023). Enhanced deep deterministic policy gradient algorithm using grey wolf optimizer for continuous control tasks. IEEE Access,11, 139771-139784. doi: 10.1109/access.2023.3341507 Suresh, A., Yüksel, M., Meder, M., Domínguez, R., & Brinkmann, W. (2025, March). Pioneering sustainable space ecosystems through intelligent robotics and collaborative effort. In Easn 2024 (p. 76). MDPI. doi: 10.3390/engproc2025090076 Svotina, V., & Cherkasova, M. (2023, March). Space debris removal - review of technologies and techniques. flexible or virtual connection between space debris and service spacecraft. Acta Astronautica,204 , 840-853. doi: 10.1016/j.actaastro.2022.09.027 UNOOSA. (2021). Guidelines for the long-term sustainability of outer space activities of the committee on the peaceful uses of outer space. https://www.unoosa.org/documents/ pdf/PromotingSpaceSustainability/Publication_Final_English_June2021.pdf. (Accessed: 2 August 2025) Vasconcelos, J., Gaggi, S., Amaral, T., Bakouche, C., Cotuna, A., & Friaças, A. (2025, January). Close-proximity operations design, analysis, and validation for non-cooperative targets with an application to the clearspace-1 mission. Aerospace,12(1), 67. doi: 10.3390/ aerospace12010067 Vela, C., Fasano, G., & Opromolla, R. (2022, December). Pose determination of passively cooperative spacecraft in close proximity using a monocular camera and aruco markers. Acta Astronautica,201 , 22-38. doi: 10.1016/j.actaastro.2022.08.024 34 Wang, W., & Chen, Z. (2025). Observer-based deep reinforcement learning for robust missile guidance and control. IEEE Access,13 , 32769-32780. doi: 10.1109/access.2025.3542129 Wilson, A. R., & Vasile, M. (2023, October). The space sustainability paradox. Journal of Cleaner Production,423 , 138869. doi: 10.1016/j.jclepro.2023.138869 Wu, C., Zhao, P., Chen, P., Ni, Z., Yue, S., Li, L., . .. Hao, G. (2025, June). Design, dynamic modelling and experimental study on a tether-net system for active debris removal. Mechanism and Machine Theory,208, 105958. doi: 10.1016/j.mechmachtheory.2025.105958 Xu, R., Zhao, Y., Li, Z., Zhu, S., Liang, Z., & Gao, Y. (2023, January). Hierarchical multiagent planning for flexible assembly of large-scale lunar facilities. Advanced Engineering Informatics,55, 101861. doi: 10.1016/j.aei.2022.101861 Zhang, Z., Wang, J., Li, Z., Wang, Y., & Zheng, J. (2025, July). Anncoder: A mti-agentbased code generation and optimization model. Symmetry,17 (7), 1087. doi: 10.3390/ sym17071087 35