Paper deep dive
SpaceMind: A Modular and Self-Evolving Embodied Vision-Language Agent Framework for Autonomous On-orbit Servicing
Aodi Wu, Haodong Han, Xubo Luo, Ruisuo Wang, Shan He, Xue Wan
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/18/2026, 1:34:19 AM
Summary
SpaceMind is a modular, self-evolving vision-language model (VLM) agent framework designed for autonomous on-orbit servicing. It features a decoupled architecture using MCP-Redis for environment-agnostic operation, dynamic skill routing, and three switchable reasoning modes (Standard, ReAct, Prospective). The framework enables agents to learn from failure and improve performance through a self-evolution mechanism that distills experience into persistent skill files without model fine-tuning.
Entities (5)
Relation Signals (4)
SpaceMind â implements â Skill Self-Evolution
confidence 100% ¡ We introduce a mechanism that enables the agent to autonomously distill operational experience
SpaceMind â utilizes â VLM
confidence 100% ¡ SpaceMind uses a VLM as a decision-control hub
SpaceMind â validatedin â UE5
confidence 95% ¡ We validate SpaceMind through 192 closed-loop runs across five satellites... a UE5 simulation
MCP-Redis â connects â SpaceMind
confidence 90% ¡ An MCP-Redis interface layer enables the same codebase to operate across simulation and physical hardware
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Autonomous on-orbit servicing demands embodied agents that perceive through visual sensors, reason about 3D spatial situations, and execute multi-phase tasks over extended horizons. We present SpaceMind, a modular and self-evolving vision-language model (VLM) agent framework that decomposes knowledge, tools, and reasoning into three independently extensible dimensions: skill modules with dynamic routing, Model Context Protocol (MCP) tools with configurable profiles, and injectable reasoning-mode skills. An MCP-Redis interface layer enables the same codebase to operate across simulation and physical hardware without modification, and a Skill Self-Evolution mechanism distills operational experience into persistent skill files without model fine-tuning. We validate SpaceMind through 192 closed-loop runs across five satellites, three task types, and two environments, a UE5 simulation and a physical laboratory, deliberately including degraded conditions to stress-test robustness. Under nominal conditions all modes achieve 90--100% navigation success; under degradation, the Prospective mode uniquely succeeds in search-and-approach tasks where other modes fail. A self-evolution study shows that the agent recovers from failure in four of six groups from a single failed episode, including complete failure to 100% success and inspection scores improving from 12 to 59 out of 100. Real-world validation confirms zero-code-modification transfer to a physical robot with 100% rendezvous success. Code: this https URL
Tags
Links
- Source: https://arxiv.org/abs/2604.14399v1
- Canonical: https://arxiv.org/abs/2604.14399v1
Trouble viewing inline? Open PDF directly â
Full Text
67,217 characters extracted from source content.
Expand or collapse full text
SpaceMind: A Modular and Self-Evolving Embodied Vision-Language Agent Framework for Autonomous On-orbit Servicing Aodi Wu 1,2 Haodong Han 1,2 Xubo Luo 1,2 Ruisuo Wang 2 Shan He 2 Xue Wan 2â 1 University of Chinese Academy of Sciences, Beijing, China 2 Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences, Beijing, China Abstract Autonomous on-orbit servicing demands embodied agents that perceive through visual sen- sors, reason about 3D spatial situations, and execute multi-phase tasks over extended hori- zons. We present SpaceMind, a modular and self-evolving vision-language model (VLM) agent framework that decomposes knowledge, tools, and reasoning into three independently exten- sible dimensions: skill modules with dynamic routing, Model Context Protocol (MCP) tools with configurable profiles, and injectable reasoning-mode skills. An MCP-Redis interface layer enables the same codebase to operate across simulation and physical hardware without modifi- cation, and a Skill Self-Evolution mechanism distills operational experience into persistent skill files without model fine-tuning. We validate SpaceMind through 192 closed-loop runs across five satellites, three task types, and two environments, a UE5 simulation and a physical lab- oratory, deliberately including degraded conditions to stress-test robustness. Under nominal conditions all modes achieve 90â100% navigation success; under degradation, the Prospective mode uniquely succeeds in search-and-approach tasks where other modes fail. A self-evolution study shows that the agent recovers from failure in four of six groups from a single failed episode, including complete failure to 100% success and inspection scores improving from 12 to 59 out of 100. Real-world validation confirms zero-code-modification transfer to a physical robot with 100% rendezvous success. Code: https://github.com/wuaodi/SpaceMind. 1 Introduction The growing population of space debris and aging satellites in low Earth orbit, which exceeded 20,000 cataloged objects as of 2023 [1], has made on-orbit servicing, assembly, and manufacturing (OSAM) an increasingly critical capability [2, 3]. Missions such as NASAâs OSAM-2 [4], ESAâs ClearSpace-1 [5], and RemoveDEBRIS [6] have demonstrated early concepts for active debris re- moval and satellite life extension. However, current operational approaches rely on pre-programmed sequences designed for known cooperative targets, limiting their ability to handle the diverse and unpredictable conditions of real on-orbit encounters. Recent studies have begun exploring vision- language models (VLMs) as spacecraft operator agents [7, 8], suggesting that large foundation models can decompose complex space tasks through natural-language reasoning. Yet these efforts â Corresponding author. Email: wuaodi20@mails.ucas.ac.cn (Aodi Wu), wanxue@csu.ac.cn (Xue Wan). 1 arXiv:2604.14399v1 [cs.RO] 15 Apr 2026 UE5 Simulation Target Robot Lab Service Service Target Perception VLM Reason Action SpaceMind Rendezvous Search and Approach Inspection ¡ Type ¡ Name ¡ Function ¡ Status ¡ Parts Modular Self-Evolving Embodied ¡ Skill Dynamic Routing ¡ Plug-and-Play Tools ¡ Switchable Reasoning ¡ Learn from Failure ¡ Fail â 100% Recovery ¡ 12 â 68 pts Boost ¡ 6-DoF Physical Space ¡ Long-Horizon Tasks ¡ Active Sensor Control 5 Satellites ¡ 174 runs 2 Satellites ¡ 18 runs Figure 1: Overview of SpaceMind. The framework operates as a VLM-based decision-control hub that perceives through visual sensors, reasons about the current situation, and issues motion and sensor-control commands. It supports three task types, namely rendezvous, search-and-approach, and inspection, across both high-fidelity UE5 simulation with 5 satellites and 174 runs and a physical laboratory with 2 satellites and 18 runs. Key capabilities include modular skill routing, plug-and- play tools, switchable reasoning modes, and experience-driven self-evolution. remain at the demonstration level, operating in simplified game environments by reading screen text rather than processing actual visual sensor data, addressing only single tasks with monolithic prompts, and lacking systematic evaluation across targets or environments. More broadly, the current AI agent research landscape [9] is dominated by digital-world applications such as web navigation, code generation, and dialogue systems, or simple physical scenarios such as tabletop manipulation and indoor navigation. Autonomous on-orbit servicing, by contrast, demands embod- ied agents that perceive through real sensors, execute 6-degree-of-freedom (DoF) motion control in a physical 3D environment, and sustain multi-phase tasks spanning dozens of decision steps, from 2 long-range search to close-range approach to fine-grained diagnosis (Fig. 1). How to build a scalable, adaptable, and sensor-driven agent framework for this domain remains an open problem. Building such a framework entails four interrelated challenges. First, there is no modular agent architecture designed for space operations. Existing space agent prototypes encode all task knowl- edge in a single monolithic prompt, which becomes unmaintainable as the number of tasks grows; simulation and real hardware typically require separate codebases, with no interface abstraction to bridge them. Second, space agents lack flexible reasoning depth and cross-step memory. Different task complexities call for different decision depths; a straightforward rendezvous may require only di- rect action selection, whereas a search-and-approach task benefits from deliberative multi-candidate evaluation. Yet current approaches offer no mechanism to switch reasoning strategies. Moreover, agents need persistent memory across decision steps to recover from target loss and dynamically adjust strategies, which existing space agent designs do not support. Third, space agents cannot improve from experience. Current designs are one-shot executors that start each episode from scratch, forgoing the operational knowledge gained in prior runs. Space operations are inherently repeated-execution scenarios where the same task types recur across different targets and conditions; an agent that accumulates experience could substantially improve long-term reliability. Traditional adaptation methods are impractical for on-orbit deployment: fine-tuning requires labeled data and gradient updates, reinforcement learning demands explicit reward function design, and in-context learning is bounded by the context window and does not persist across sessions. Fourth, the design decisions behind space agent architectures lack systematic empirical validation. Prior studies [7, 8] report results on a single task with a handful of runs, providing no evidence on whether modular architectures scale, which reasoning mode to select, whether self-evolution is effective, or whether the same architecture can transfer across environments. To address these challenges, we present SpaceMind, a modular and self-evolving embodied VLM- agent framework for autonomous on-orbit servicing (Fig. 2). SpaceMind uses a VLM as a decision- control hub that perceives the environment through visual sensors, reasons about the current situ- ation, invokes specialized tools, and issues motion commands. Its design comprises four progressive layers. To address the modularity challenge, SpaceMind decomposes agent knowledge into indepen- dent skill modules with LLM-based dynamic routing, exposes all executable tools through the Model Context Protocol (MCP) [10] with configurable tool profiles, and switches reasoning modes via in- jectable mode skills. These three orthogonal dimensions are managed through a single declarative configuration. A Redis message bus decouples the agent from the physical backend, allowing the same codebase to operate across simulation and real hardware without modification. To address the reasoning challenge, SpaceMind hosts three reasoning modes within the same architecture: Standard for direct decision, ReAct [11] for iterative thoughtâactionâobservation loops, and Prospective for multi-candidate prediction and selection. These modes are complemented by a hierarchical mem- ory system that maintains recent-step detail and long-term compressed summaries. To address the experience accumulation challenge, a Skill Self-Evolution mechanism enables the agent to reflect on each completed episode, distill operational experience into structured learned skill files, and auto- matically inject them into subsequent runs, thereby accumulating persistent, auditable knowledge without modifying model weights. Finally, to address the validation challenge, we conduct com- prehensive closed-loop experiments totaling 192 runs across five satellites, three task types, three reasoning modes, three initial conditions, and two environments, a high-fidelity UE5 simulation and a physical laboratory. The experiments yield several actionable findings. A tool ablation study of 9 runs shows that removing LiDAR range sensing causes navigation failure, validating the modular tool configuration mechanism. A systematic reasoning evaluation spanning 135 runs under nominal and degraded ini- tial conditions such as underexposure and positional offset reveals that all modes achieve 90â100% 3 navigation success under nominal conditions but exhibit sharply divergent degradation patterns when rendezvous and search-and-approach are examined separately: Standard maintains 100% ren- dezvous success even under degraded conditions but its search success drops to zero. Prospective is the only mode that maintains search success under degradation, while ReActâs rapid termination strategy proves advantageous for inspection under poor visibility. These findings yield concrete mode-selection guidelines for deployment. A self-evolution study covering 30 runs across 6 groups demonstrates that four of six groups achieve significant positive gains. Most notably, one group recovers from complete failure to 100% success and another raises its inspection score from 12 to 59 out of 100, in both cases by generating learned skills from a single failed episode. Real-world vali- dation over 18 runs confirms that the identical agent codebase, skill definitions, and tool signatures operate a physical mobile robot with zero code modification, achieving 100% rendezvous success and demonstrating emergent scale adaptation and behavioral consistency across environments. The contributions of this paper are summarized as follows: 1. Modular agent architecture for on-orbit servicing. We propose SpaceMind, a VLM- agent framework tailored for autonomous spacecraft operations that decomposes agent knowl- edge, tools, and reasoning into three independently extensible dimensions, namely skill mod- ules with dynamic routing, MCP-based tools with configurable profiles, and injectable reasoning- mode skills. An environment-agnostic interface layer enables the identical codebase to operate across a high-fidelity space simulation and a physical laboratory without modification. 2. Multi-mode reasoning for multi-phase space tasks. We implement three switchable reasoning modes, Standard, ReAct, and Prospective, within a unified architecture, providing configurable cognitive depth for tasks ranging from straightforward rendezvous to long-horizon search-and-approach under degraded conditions. A hierarchical memory system supports cross-step context reasoning for target-loss recovery and strategy adjustment over extended episodes. 3. Skill Self-Evolution for autonomous on-orbit improvement. We introduce a mech- anism that enables the agent to autonomously distill operational experience into persistent, structured skill files after each servicing episode, without model fine-tuning or reward function design, achieving significant recovery from failure in four of six experimental groups. 4. Comprehensive closed-loop validation. We conduct the first systematic evaluation of a VLM-based space agent, spanning 192 runs across five satellites, three tasks, three rea- soning modes, and two complementary environments, providing empirical grounding for each architectural design decision. 2 Related Work 2.1 Autonomous Spacecraft Operations and VLM-based Space Agents Autonomous on-orbit servicing has traditionally relied on pre-programmed sequences for known cooperative targets, with perception handled by dedicated pipelines for spacecraft detection [12], part-level semantic segmentation [13, 14], and relative pose estimation [15â17]. These methods have advanced substantially in recent years, as surveyed by Pauly et al. [18], yet they operate as isolated modules that must be manually integrated into mission-specific control flows, limiting adaptability to new targets or tasks. A recent line of work has begun exploring large language and vision-language models as high- level spacecraft operators. Rodriguez-Fernandez et al. [7] demonstrated that language models can 4 decompose spacecraft control tasks through natural-language reasoning, and Carrasco et al. [8] ex- tended this to visual-language models for operator-level decision-making. However, these studies share three critical limitations. They address single tasks with monolithic prompts and no modular framework, providing no path to multi-task scalability. They operate in simplified game environ- ments by reading screen text and user-interface elements rather than processing actual visual sensor data, making them text-based agents rather than truly embodied, sensor-driven systems. They also report results on a small number of runs with no cross-environment validation. Separately, Foutter et al. [19] explored adapting a foundation model for Mars surface navigation, but without closed-loop control or multi-task evaluation. In contrast, SpaceMind processes raw RGB camera and LiDAR data through a modular architecture, operates in both high-fidelity UE5 simulation and a physical laboratory, and is validated across 192 closed-loop runs spanning five satellite targets and three task types. 2.2 LLM/VLM-based Embodied Agents and Tool Integration The broader AI agent community has made rapid progress in building autonomous systems around large foundation models. Wang et al. [9] survey the emerging paradigm of LLM-based agents com- prising a language model core, memory, tool use, and planning capabilities. On the tool-integration front, Qin et al. [20] formalize the framework of tool learning with foundation models, while the Model Context Protocol (MCP) [10] provides a standardized interface for connecting agents to ex- ternal tools and data sources. Chain-of-thought prompting [21] showed that explicit intermediate reasoning steps substantially improve LLM decision quality, and Tree of Thoughts [22] extended this idea to deliberate search over multiple reasoning branches. ReAct [11] further introduced the influential pattern of interleaving reasoning traces with tool actions, enabling agents to ground their decisions in environmental feedback. In embodied settings, PaLM-E [23] demonstrated multimodal reasoning grounded in robotic observations, SayCan [24] and Inner Monologue [25] enabled language-guided tool use with environ- mental feedback, and vision-language-action (VLA) models such as RT-2 [26] and OpenVLA [27] have demonstrated end-to-end visuomotor control for robotic manipulation. GPT-Driver [28] ap- plied LLM-based planning to autonomous driving. These approaches have primarily targeted ter- restrial environments with relatively short task horizons and low-dimensional action spaces (e.g., tabletop grasping, lane following). Three gaps remain when extending this paradigm to space oper- ations. None of these systems addresses 6-DoF long-horizon tasks in unstructured 3D environments. Most adopt a single, fixed reasoning strategy, with no systematic comparison of alternative reason- ing modes. Furthermore, tool integration is typically hard-coded for a specific environment, lacking the abstraction layer needed for cross-environment portability. SpaceMind addresses these gaps through an MCP-Redis interface architecture that decouples the agent from the physical backend, a modular skill layer with LLM-based dynamic routing that replaces monolithic prompts, and three switchable reasoning modes evaluated systematically across 135 runs. 2.3 Experience-driven Self-Improvement for Agents Enabling agents to improve autonomously from deployment experience is a long-standing goal. Existing adaptation paradigms each carry significant limitations in the context of space operations. Fine-tuning requires labeled training data and gradient computation, both impractical on orbit. Standard in-context learning is bounded by the context window and does not persist across sessions. Retrieval-augmented generation (RAG) retrieves existing documents but cannot synthesize new operational knowledge from experience. Reinforcement learning demands explicit reward function 5 design and is sample-inefficient in physical environments. Several recent works have explored lighter-weight self-improvement mechanisms for language agents. Reflexion [29] introduced verbal reinforcement learning, where an agent reflects on its failures and stores textual feedback for subsequent trials. However, the reflections remain in the context window and are lost across sessions. ExpeL [30] extracts reusable âexperiencesâ from trajectories and stores them as persistent insights, yet the extracted knowledge remains natural-language summaries without structured format or quality verification. Voyager [31] builds a persistent skill library by generating executable code functions in Minecraft, representing the closest conceptual precedent to our approach; yet it operates in a digital sandbox with deterministic physics, and its skills are code snippets rather than structured operational knowledge that can be audited or composed with existing domain expertise. SpaceMindâs Skill Self-Evolution mechanism occupies a distinct niche: it generates persistent, structured, auditable operational knowledge from real deployment experience in physical-world space tasks. The learned skills share the same format as hand-authored skills, enabling seamless integration with the modular skill architecture and automatic routing by the skill gateway. Unlike Reflexion, the knowledge persists across sessions; unlike Voyager, it targets physical 3D operations with safety-critical constraints and includes a quality gate with fingerprint deduplication, safety phrase filtering, and task-scope binding to prevent cross-task contamination. Experimental valida- tion across 30 runs in six groups demonstrates that four groups achieve significant improvement, with the mechanism proving particularly effective at recovering from failure and iteratively refining perception strategies. 3 Methodology 3.1 System Overview SpaceMind is a closed-loop embodied agent framework in which a vision-language model (VLM) serves as a central decision-control hub. Fig. 2 illustrates the overall architecture. At each decision step t, the agent receives an observation o t consisting of an RGB image and an optional LiDAR range summary from the environment. A system prompt is assembled by composing dynamically routed skill modules, reasoning-mode instructions, and a hierarchical memory context summarizing prior steps. The VLM processes this prompt together with the visual observation to produce an action a t in the form of a tool call (e.g., set_position, set_attitude, set_exposure), which is dispatched through the Model Context Protocol (MCP) [10] and relayed to the physical environment via a Redis message bus. The environment executes the action, produces a new observation o t+1 , and the loop continues until a termination condition is reached. Formally, the decision process at each step is a t = Ď VLM o t , S t , M t ; θ ,(1) whereS t denotes the assembled skill prompt (determined by the skill gatewayâs routing result), M t the hierarchical memory context, and θ the frozen VLM parameters. The environment transition follows o t+1 =E o t , a t ,(2) where E abstracts the physical backend (UE5 simulation or laboratory hardware), decoupled from the agent through the Redis message bus. The architecture is organized around four design layers, each addressing a specific challenge identified in Section 1. The Skill Layer with dynamic routing addresses the modularity challenge by decomposing agent knowledge into independently composable skill modules (Sec. 3.2). The 6 Redis Message Bus VLM Decision Core Direct decision Standard Think â Act â Observe ReAct Planâ Score â Select Prospective Reasoning Modes (Switchable vis mode skill injection) Cognitive Depth Recent Steps + Long-Term Summary Memory Manager Task â dynamic selection skills Skill Gateway (LM Router) always-on Core route 1 Task 0-2 Helper byconfig Mode Self-evolved Learned Perception Vision analysis Control Motion Exposure Auxiliary Knowledge Code Execution Skill Layer MCP Tool Layer RGB + Lidar | 6DoF UE5/AirSim RGB | 3DoF planar AGV Pro tool callimage / lidarsystem prompt Self-Evolution running logs Summarize trajectory LLM analyzes experience Episode Reflection Load into skill layer next episode Auto Injection Safety check Validation Quality Gate Same format ad base skills Learned SKILL.md Figure 2: SpaceMind architecture. The framework comprises four layers. The Skill Layer decomposes agent knowledge into modular skill definitions with LLM-based dynamic rout- ing (Sec. 3.2). The VLM Decision Core hosts three switchable reasoning modes with hierarchical memory (Sec. 3.3). The MCP Tool Layer and Redis Message Bus provide an environment-agnostic interface that enables the same codebase to operate across UE5 simulation and a physical labora- tory (Sec. 3.4). A side loop implements Skill Self-Evolution, which distills episode experience into learned skill files and injects them into subsequent runs (Sec. 3.5). Reasoning and Memory layer addresses the cognitive-depth challenge by hosting three switchable reasoning modes and a hierarchical memory system (Sec. 3.3). The Interface and Communication Architecture addresses cross-environment portability through MCP tool abstraction and a Redis message bus (Sec. 3.4). The Skill Self-Evolution mechanism addresses experience accumulation by generating persistent learned skills from deployment episodes (Sec. 3.5). Algorithm 1 summarizes the complete execution flow of a single episode, providing a unified view of the four layers before each is detailed in the following subsections. 3.2 Skill Layer and Dynamic Routing Existing VLM-based agent approaches for space operations encode all task knowledge in a sin- gle monolithic system prompt, which creates three problems as the operational scope expands: the prompt becomes unmanageable as more tasks are added, shared strategies (e.g., distance-dependent step sizing, target-loss recovery) cannot be reused across tasks, and experience accumulated during operation has no structured container for storage. To address these limitations, SpaceMind decom- poses agent knowledge into independent skill modules and routes them dynamically based on the task context. Each skill is a self-contained knowledge unit consisting of a structured header and a natural- 7 Algorithm 1 SpaceMind Episode Execution Input: task description d, reasoning mode m, tool profile P, skill catalog C, max steps T Output: trajectory Ď, outcome r // Phase 1: Initialization 1: (s task , S helper )â SkillGateway(d, P, m, C) 2: S â S core âŞs task ⪠S helper ⪠ModeSkill(m)⪠LearnedSkills(d) 3: M 0 ââ ; o 0 âE.Reset() // Phase 2: Decision Loop 4: for t = 0, 1,...,Tâ1 do 5: if m = Standard then 6:a t â Ď VLM (o t , S, M t ; θ)⡠Eq. (1) 7: else if m = ReAct then 8:a t â ReActLoop(o t , S, M t , R=3) 9: else if m = Prospective then 10: c 1 ,c 2 ,c 3 â VLMPlan(o t , S, M t ) 11:a t â VLMSelect(c 1 ,c 2 ,c 3 , o t ) 12: end if 13: resultâ MCP.Execute(a t )⡠tool call via Redis 14: o t+1 âE.Observe() 15: M t+1 â UpdateMemory(M t , a t , result, o t+1 ) 16: if IsTerminated(a t , result) then break 17: end if 18: end for // Phase 3: Post-Episode Self-Evolution 19: Ď â CollectTrajectory(); r â Evaluate(Ď) 20: mutâ VLMReflect(Summarize(Ď,r), S, history) 21: if mut̸= NoChange and QualityGate(mut) then 22: S learned â S learned ⪠Materialize(mut) 23: end if 24: return Ď, r language body. The header declares the skillâs name, category, a routing summary describing its applicability, and routing keywords for catalog matching. The body contains the operational instructions that are injected verbatim into the VLMâs system prompt. Skills are organized into a three-tier taxonomy: ⢠Core skills are always active and provide foundational operational knowledge shared across all tasks, including coordinate conventions, safety constraints, and general action strategies. ⢠Task skills encode the primary strategy for a specific task type (e.g., rendezvous, search-and- approach, inspection). Exactly one task skill is selected per episode. ⢠Helper skills provide auxiliary strategies that complement the primary task skill, such as distance-dependent step-size decay, target-loss recovery procedures, or multi-modal perception workflows. Zero to two helper skills are selected per episode. A skill gateway performs dynamic routing at the start of each episode. The gateway takes the current task description, active tool profile, and reasoning mode as input, and issues a lightweight LLM call over a catalog of available task and helper skills. The LLM selects one primary task 8 skill and up to two complementary helper skills, along with a natural-language justification for its selection. If the routing call fails (e.g., due to malformed output), the gateway falls back to a default skill combination declared in the framework configuration, ensuring robustness. In addition to task and helper skills, mode skills provide reasoning-mode-specific instructions (detailed in Sec. 3.3). These are injected based on the active reasoning mode and are fully decoupled from task skills, so that any combination of task and reasoning mode can be composed without interference. Skills generated by the self-evolution mechanism (Sec. 3.5) share the same format as hand-authored skills, enabling seamless injection through the same routing pipeline. The final system prompt is assembled by concatenating the active core skills, the routed task and helper skills, the mode-specific skills, and any applicable learned skills, in that order. All skill assignments, tool profiles, reasoning modes, and gateway parameters are managed through a single declarative configuration file, achieving four-dimensional decoupling: any combination of task, reasoning mode, tool set, and skill composition can be specified without code changes. 3.3 Reasoning Modes and Memory Different operational scenarios demand different levels of cognitive depth. A straightforward ren- dezvous with a visible target may require only direct action selection, whereas navigating under poor visibility or searching for an out-of-view target benefits from deliberative multi-step reasoning. SpaceMind addresses this by hosting three reasoning modes within the same architecture, forming a cognitive-depth gradient that can be selected per episode. Standard mode The VLM receives the current observation, the assembled system prompt, and the list of available tools, and directly produces a single tool call as its action. If the selected tool is a perception operation (e.g., image analysis or segmentation), its result can trigger a follow-up VLM call to select a subsequent motion action within the same step. This mode incurs the lowest computational overhead and is suited for tasks where the correct action can be determined from the current observation alone. ReAct mode Following the ReAct paradigm [11], each decision step allows up to three internal rounds of ThoughtâActionâObservation. The agent first generates an explicit reasoning trace, then selects and executes a tool, and observes the result before deciding whether to continue reasoning or commit to a motion action. The inner loop terminates when a motion or termination tool is called. This mode is activated by injecting a mode skill that instructs the VLM to interleave reasoning with tool use, and is suited for tasks requiring information gathering before committing to an action. Prospective mode Inspired by deliberative search strategies such as Tree of Thoughts [22], the agent employs a two-phase deliberation process for high-uncertainty scenarios. In the planning phase, the VLM generates three candidate actions, each accompanied by a predicted outcome and a risk assessment. In the selection phase, a second VLM call evaluates the candidates and selects the one with the most favorable predicted outcome. This mode is activated by injecting two com- plementary mode skills, one for planning and one for selection, and provides the highest cognitive depth, enabling the agent to reason about action consequences before committing. If the planning phase produces malformed output, the system gracefully degrades to a direct selection call. All three modes are activated exclusively through mode skill injection and are fully decoupled from task-specific skills. Switching between modes requires only a configuration change, with no modification to the agentâs decision loop or task knowledge. 9 Hierarchical memory To support cross-step reasoning, SpaceMind maintains a two-tier memory system. The recent memory retains detailed records of the most recent N steps, including the agentâs analysis, tool selections, arguments, and results. The long-term memory compresses older steps into a natural-language summary that captures cumulative trajectory information. Both tiers are concatenated and injected as contextual input to each VLM call, enabling the agent to detect and recover from target loss, avoid repeating failed strategies, and track progress toward task completion across extended episodes. 3.4 Interface and Communication Architecture A key design goal of SpaceMind is that the same agent codebase operates across different physical environments without modification. Achieving this requires decoupling the agentâs decision logic from the specifics of any particular sensor or actuator backend. SpaceMind implements this decou- pling through two abstraction layers: an agent-tool protocol and a tool-environment communication bus. MCP tool protocol All executable capabilities are exposed to the agent as tools through the Model Context Protocol (MCP) [10], a standardized interface for connecting language models to external functions. The agent invokes tools by name and argument via a unified calling convention, without knowledge of how each tool is implemented. Tools are organized into four functional categories: ⢠Perception tools analyze the current visual input, including brightness assessment, part-level segmentation, region cropping, and zoom. ⢠Control tools issue motion and sensor commands such as translational and rotational motion, camera exposure adjustment, and task termination. ⢠Knowledge tools provide access to a domain knowledge base describing spacecraft character- istics. ⢠Auxiliary tools support optional capabilities such as code execution for numerical computation. A tool profile mechanism controls which tools are visible to the agent for a given experiment. Each profile specifies an allowed-tool whitelist; at runtime, only tools appearing in both the MCP serverâs full catalog and the active profileâs whitelist are presented to the VLM. This enables controlled ablation studies (e.g., removing range sensing to test vision-only operation) and environment-specific tool sets (e.g., a minimal profile for the physical laboratory) through configuration alone. Two design decisions merit discussion. First, translational and rotational motion are exposed as separate tools rather than a single 6-DoF command, preventing the VLM from producing coupled translation-rotation actions that lead to unintended spiral trajectories, a failure mode observed in early development. Second, camera exposure adjustment is exposed as an explicit tool, enabling the agent to actively improve its own perception quality when it detects underexposure or overexposure, rather than passively accepting degraded imagery. Redis message bus Between the MCP tool implementations and the physical environment, a Redis publish-subscribe message bus provides the second abstraction layer. A unified naming contract defines the communication channels: sensor data flows upstream through designated keys (e.g., latest RGB image, latest LiDAR summary) and image-arrival notifications, while control commands flow downstream through corresponding command topics (e.g., pose-change increments, 10 exposure adjustments). In the UE5 simulation, an AirSim [32] bridge process writes sensor data to Redis and consumes motion commands. In the physical laboratory, a ROS2 sensor bridge publishes camera frames to the same Redis keys, and a separate motion executor subscribes to the same command topics to drive the mobile robot. Both backends populate and consume identical Redis channels; the agentâs decision loop contains no environment-conditional logic. This two-layer abstraction yields a concrete portability guarantee: the same agent decision loop, MCP tool server, and communication contract operate in both environments with zero code modification. Extending SpaceMind to a new environment backend requires only implementing the Redis read/write endpoints for that backendâs sensors and actuators. 3.5 Skill Self-Evolution Mechanism Space on-orbit servicing is inherently a repeated-execution domain: the same task types recur across different targets and conditions. An agent that discards all operational experience after each episode forfeits the opportunity to improve from its own successes and failures. Traditional adaptation methods are poorly suited to this setting: model fine-tuning requires labeled training data and gradient computation, reinforcement learning demands explicit reward function design, in-context learning is bounded by the context window and does not persist across sessions, and retrieval- augmented generation can only retrieve existing documents rather than synthesize new knowledge. SpaceMind introduces a Skill Self-Evolution mechanism that enables the agent to autonomously generate, validate, and accumulate operational knowledge across episodes without modifying model weights. The mechanism operates as an outer loop around the standard decision cycle (Fig. 2, right side). After each episode concludes, the following pipeline executes: Episode summarization The system records the complete tool-call trajectory, success or failure outcome, termination reason, and a structured summary of the agentâs movement and perception history during the episode. Experience reflection The episode summary, together with summaries from recent prior episodes and the currently active skill definitions, is provided to the VLM in a dedicated reflection call. The VLM analyzes what worked, what failed, and why, and produces a structured mutation decision specifying one of five actions: create a new learned skill capturing a previously unknown operational pattern; overlay an existing skill with refined or supplementary rules; rewrite an existing skill when evidence warrants a fundamental revision; disable a learned skill that has been found harmful; or no_change when the current experience does not justify a modification. Quality gate Every proposed mutation passes through a multi-stage validation gate before being applied. The gate enforces four constraints: ⢠A safety-phrase blacklist rejects mutations containing instructions that could override safety rules. ⢠A fingerprint deduplication check prevents semantically redundant skills from accumulating. ⢠A task-scope binding ensures that skills learned in one task context cannot contaminate an- other taskâs knowledge base. ⢠A parent-skill validation confirms that overlay and rewrite operations target skills that actually exist. 11 Mutations that fail any constraint are discarded, and an audit log records the rejection reason. Learned skill generation Mutations that pass the quality gate are materialized as new skill files sharing the same format as hand-authored skills, ensuring compatibility with the existing skill routing infrastructure. Each learned skill contains five structured sections: Intent specifies what the skill aims to achieve, Trigger defines the conditions under which the skill should activate, Rule encodes the operational instructions, Constraints delineates safety boundaries, and Evidence records the episode observations that motivated the skill. Additional metadata records the skillâs provenance, version, originating episode, and applicable task scope. Automatic injection At the start of each subsequent episode, the skill runtime loads all learned skills alongside the base skill set. A retrieval function matches learned skills to the current task and reasoning mode based on their declared scope and trigger conditions, selecting the top-k most relevant learned skills for injection into the system prompt alongside the dynamically routed base skills. This design has four properties that distinguish it from prior self-improvement approaches. First, learned skills are naturally compatible with the modular skill architecture (Sec. 3.2): they use the same data format, pass through the same routing pipeline, and compose with base skills without special handling. Second, the knowledge is persistent and auditable: each mutation is recorded with full provenance, and the skill files can be inspected, edited, or reverted by human operators. Third, the evolution is progressive: skills accumulate incrementally through versioned overlays rather than wholesale replacement, allowing gradual refinement. Fourth, the quality gate provides a safety guarantee: even if the VLMâs reflection produces an unsafe or redundant suggestion, the hard constraints prevent it from entering the active skill set. 4 Experimental Validation We evaluate SpaceMind through 192 closed-loop runs organized into four experimental campaigns, conducted across a high-fidelity Unreal Engine 5 (UE5) simulation and a physical laboratory envi- ronment. Table 1 summarizes the campaigns. The tool ablation campaign validates the modular tool configuration mechanism. The reasoning evaluation campaign provides a systematic comparison of reasoning modes across five satellites, three tasks, and three initial conditions. The self-evolution campaign tests autonomous skill improvement through iterative execution. The real-world campaign validates cross-environment portability with zero code modification. Table 1: Overview of experimental campaigns. CampaignEnvRuns Tool Ablation UE59 Reasoning Eval UE5135 Self-Evolution UE530 Real-WorldAGV Pro 18 Total192 12 4.1 Experimental Setup Simulation environment The simulation is built on UE5 with the AirSim plugin [32], extending the SpaceSense-Bench platform [33] to provide high-fidelity optical rendering and multi-sensor sim- ulation. Five satellite models of varying geometry and scale are placed in the scene: CAPSTONE, IBEX, BioSentinel, New Horizons, and Huygens. The agent perceives the environment through an RGB camera and a LiDAR range sensor, and controls its 6-DoF position and attitude via MCP [10] tool calls relayed over a Redis message bus. Laboratory environment The physical testbed consists of a myAGV Pro omnidirectional mobile robot with Mecanum wheels and a 40 kg payload, equipped with an Orbbec Gemini 2 stereo RGB camera at 1920Ă 1080 resolution. The robot runs ROS2 Humble and communicates with the same agent decision loop through a Redis bridge. Two 1-meter-diameter 3D-printed satellite mockups, CAPSTONE and Artemis, serve as targets. Motion is constrained to a 3-DoF planar workspace. Tasks Three task types of increasing difficulty are defined in Table 2. Table 2: Task definitions and evaluation metrics. TaskAbbr. DescriptionMetrics RendezvousRndz. Target visible; approach to âź2 mSucc, Dist, Steps Search-and-Approach Search Target not visible; find then approach to âź2 m Succ, Dist, Steps InspectionInsp. Close-range 5-dim structured reportScore (0â100) Initial conditions Three initial conditions are used in the UE5 simulation to evaluate robustness (Table 3). All three tasks are tested under each condition. Table 3: Initial condition definitions for UE5 simulation. Cond. PositionExposureDifficulty C1NominalNominalBaseline C2Far lateral offset Underexposure Degraded C3Reverse offsetOverexposure Degraded VLM backbone All experiments use the open-source Qwen3-VL-235B-A22B model [34] as the VLM decision core, with temperature ⤠0.3 to promote output stability across all campaigns. 4.2 Tool Configuration Ablation To validate the modular tool configuration mechanism, the tool ablation campaign compares three tool profiles under fixed conditions, specifically Standard mode on CAPSTONE with C1. Each profile exposes a different subset of MCP tools to the agent: Vision-Only provides camera perception only, Hybrid-Nav adds LiDAR range sensing, and Hybrid-Nav-Code further adds runtime code execution. Table 4 presents the results. Without LiDAR range measurements, the Vision-Only profile fails both navigation tasks due to distance estimation errors that cause overshoot. Hybrid-Nav succeeds on all tasks with the fewest steps and highest inspection score, while adding runtime code execution does not improve 13 Table 4: Tool profile ablation results over 9 runs. Rndz.SearchNav. Pass Insp. Tool ProfileDist (m) Steps Dist (m) StepsRateScore Vision-OnlyFAILFAIL0/265 Hybrid-Nav2.00692.277162/265 Hybrid-Nav-Code 2.048132.279152/250 performance, as the agent never invokes the code-execution tool in practice. Based on these re- sults, Hybrid-Nav is adopted as the default for all subsequent experiments, demonstrating that SpaceMindâs tool profile mechanism allows declarative capability configuration. 4.3 Multi-Satellite Reasoning Mode Evaluation The reasoning evaluation campaign provides a systematic comparison across 5 satellites, 3 tasks, 3 reasoning modes (Standard, ReAct [11], Prospective), and 3 initial conditions, totaling 135 runs. 4.3.1 Results Table 5 consolidates navigation and inspection results across all modes and conditions. Fig. 3 visualizes the navigation pass rate degradation from C1 to C3. Table 5: Reasoning mode evaluation results across conditions, 135 runs in total. Navigation reports pass count per 5 runs for each task type; inspection reports average score per 5 runs. Best values in each column are bold. Rndz. (pass / 5)Search (pass / 5)Nav. TotalInsp. (avg score) ModeC1 C2 C3SubC1 C2 C3SubPassDist (m) C1C2C3 Avg Standard5 5 3 13 (87%) 4 0 0 4 (27%) 17/30 (57%) 2.143 40.2 41.0 36.2 39.1 ReAct5 1 28 (53%) 5 0 0 5 (33%) 13/30 (43%) 2.177 38.6 40.8 42.8 40.7 Prospective 5 3 2 10 (67%) 4 1 2 7 (47%) 17/30 (57%) 2.079 36.6 22.6 29.8 29.7 4.3.2 Analysis The results reveal that no single reasoning mode dominates, and the Rndz./Search split exposes fundamentally different degradation patterns. All three modes achieve 5/5 rendezvous success under C1 where the target is visible and conditions are nominal. Standard maintains perfect rendezvous even under C2 at 5/5, achieving the highest rendezvous subtotal of 13/15 at 87%. However, its search performance collapses entirely under degradation: 0/5 at both C2 and C3, yielding only 4/15 or 27%. ReAct shows a similar pattern, achieving perfect 5/5 rendezvous and search at C1 but dropping to 1/5 rendezvous and 0/5 search under C2, with most failures reaching the 50-step timeout as the ThoughtâActionâObservation loop oscillates when the target is far from the expected position. Prospective is the only mode that achieves any search success under degraded conditions: 1/5 at C2 and 2/5 at C3, yielding a search subtotal of 7/15 at 47%, nearly double Standardâs 27%. Its candidate-prediction mechanism avoids blind exploration by evaluating multiple movement hy- potheses before committing, yielding the most precise terminal distance of 2.079 m. This advantage 14 C1C2C3 Initial Condition 0 20 40 60 80 100 Navigation Pass Rate (%) Standard ReAct Prospective Figure 3: Navigation pass rate across initial conditions. ReAct degrades sharply from C1 to C2, while Standard and Prospective maintain higher robustness. is specific to search-and-approach tasks where the target must first be found; for rendezvous where the target is already visible, Standard suffices and additional reasoning overhead provides no benefit. For inspection, a counter-intuitive pattern emerges: ReAct achieves the highest score under C3 at 42.8, surpassing Standard at 36.2 and Prospective at 29.8. ReActâs rapid termination strategy of 1â3 steps limits hallucination accumulation, whereas Prospectiveâs two-stage reasoning amplifies VLM hallucination risk under degraded visual conditions. These findings validate that SpaceMindâs modular reasoning architecture can host fundamentally different strategies, and the systematic comparison provides actionable mode-selection guidance for deployment. 4.3.3 Skill Routing Validation To validate the dynamic skill routing mechanism, we extract the routing decisions from all 135 runs. The skill gateway correctly routes each task to its intended skill set: rendezvous to approach + distance skills; search-and-approach to search + target-recovery + distance skills; and inspection to inspection + perception skills. The routing achieves 100% agreement with the intended assignments across all runs without manual intervention, confirming that the LLM-based skill gateway can reliably compose task-appropriate skill sets from the modular catalog. 4.4 Skill Self-Evolution The self-evolution campaign evaluates whether the skill self-evolution mechanism can accumulate useful operational knowledge through repeated execution. We select six groups spanning three satellites and two task types, prioritizing combinations where the reasoning evaluation baseline failed or scored low. Each group undergoes five consecutive rounds with a shared evolution workspace, using Standard mode and the Hybrid-Nav tool profile. After each round, the evolution runtime reflects on the episode and may generate learned skill files that are automatically injected into subsequent rounds. Table 6 summarizes the results and Fig. 4 shows the per-round learning curves. The most striking result is the mechanismâs ability to recover from failure. The two groups 15 Table 6: Skill self-evolution results over 30 runs across 6 groups. BL = baseline from the reasoning evaluation campaign. BaselineBest After Evol. SatelliteTask Cond Result Detail ResultDetail Verdict CAPSTONE Search C1OK 2.597 m0/5â No benefit CAPSTONE Insp.C250 ptâ 65 pt+30% Positive New Horizons Search C1FAIL10 st3/52.102 m Positive New Horizons Rndz.C3FAIL20 st 5/57 st Strong + HuygensRndz.C2OK 2.209 m3/52.169 m No benefit HuygensInsp.C112 ptâ 68 pt+467% Strong + BLR1R2R3R4R5 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Success Huygens Rndz. C2 BLR1R2R3R4R5 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Success CAPSTONE Search C1 BLR1R2R3R4R5 0 20 40 60 80 Score CAPSTONE Insp. C2 BLR1R2R3R4R5 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Success New Horizons Rndz. C3 BLR1R2R3R4R5 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Success New Horizons Search C1 BLR1R2R3R4R5 0 20 40 60 80 Score Huygens Insp. C1 10 20 30 40 50 Steps 7.5 10.0 12.5 15.0 17.5 20.0 22.5 Steps 8 10 12 14 16 18 20 Steps 9 10 11 12 13 14 15 Steps Figure 4: Skill self-evolution learning curves for six groups, organized by task type in columns for Rndz., Search, and Insp. respectively. For navigation groups, bars show success and the dashed line shows step count; for inspection groups, the solid line shows score and the dashed horizontal line marks the baseline level. Green indicates clear positive benefit; gray indicates no significant improvement. that fail completely at baseline, New Horizons Search C1 and New Horizons Rndz. C3, achieve 60% and 100% success rates respectively after self-evolution. In New Horizons Rndz. C3, the step count converges steadily from 20 to 7 across five rounds, indicating that the automatically generated learned skill encodes an effective distance-dependent step-size reduction strategy. Three of the six groups show significant improvement by Round 1 or 2, demonstrating that the reflection mechanism can extract actionable knowledge from a single failed episode. For inspection tasks, both groups show consistent gains through the overlay mutation mecha- nism: CAPSTONE Insp. C2 improves from 50 to 65 points and Huygens Insp. C1 from 12 to an average of 59.4 points. The incremental overlay refinement of perception strategies proves particu- 16 larly effective for this task type. However, the two groups where the baseline already succeeds, CAPSTONE Search C1 and Huygens Rndz. C2, do not benefit from self-evolution, revealing that the mechanismâs primary value lies in error correction and strategy discovery rather than optimizing already-successful be- havior. Across all groups, seven learned skills are generated in the standard skill-file format and correctly routed by the skill gateway in subsequent rounds, confirming compatibility between the self-evolution mechanism and the modular architecture. 4.5 Real-World Validation The real-world campaign validates that the same architecture, including the identical agent loop, MCP tool signatures, and Redis communication contracts, can operate a physical robot with zero code modification; only the physical backend is swapped to a ROS2 sensor bridge and motor con- troller. Table 7 presents the results across two satellite mockups, three tasks, and three repetitions per combination, totaling 18 runs. Table 7: Real-world validation results over 18 runs. Rndz.SearchInsp. SatellitePass Runs Rate Pass Runs Rate Avg Score Runs CAPSTONE 33100%2367%41.03 Artemis33100%2367%34.03 Total66 100% 46 67%36.86 The agent achieves a perfect 6/6 pass rate on rendezvous and 4/6 on search-and-approach tasks. Three observations are specific to the physical environment. First, the agent autonomously adapts its step size from 1â2 m per step in simulation to 0.1â0.2 m per step in the laboratory, demonstrating emergent scale awareness without explicit re-configuration. Second, the target-recovery skillâs yaw Âą90 ⌠sweep pattern transfers identically from simulation, confirming behavioral consistency. Third, one inspection run results in a collision due to cumulative small steps over 15 iterations, highlighting the need for explicit safety margins in close-proximity operations. Fig. 5 illustrates a representative search-and-approach episode from the third-person perspective. These results validate the interface and communication architecture and demonstrate that SpaceMindâs modular design generalizes across environments. 4.6 Qualitative Analysis We select four representative episodes that illustrate emergent behaviors enabled by SpaceMindâs design. Fig. 6 shows keyframe images from each episode. Example 1: Lab target-lost recovery During a CAPSTONE search task in the laboratory, the target is not visible from the initial pose (Fig. 6a). Guided by the target-recovery skill, the agent executes a bounded yaw sweep of six 15 ⌠right turns followed by nine 15 ⌠left turns before detecting the target and switching to approach mode. The episode spans 42 steps, terminating at 1.13 m, demonstrating the interplay between modular skill knowledge and the hierarchical memory system. 17 Step 1 Initial pose; target not in field of view Step 4 Target detected after yaw sweep search Step 5 Heading aligned toward the target Step 8 Approaching the target at close range Step 12 Lateral drift occurred during approach Step 14 Re-aligned and halted at 0.8 m (success) Figure 5: Third-person view of a real-world search-and-approach episode on CAPSTONE over 14 steps. The agent searches for the target via yaw sweep in Steps 1â4, aligns and approaches in Steps 5â8, recovers from lateral drift at Step 12, and terminates at 0.8 m from the target surface at Step 14. Example 2: Active exposure adjustment Under the underexposed C2 condition, the ini- tial camera image is nearly black. The agent invokes a brightness diagnostic tool, then calls the exposure-control tool to increase the camera exposure. After adjustment, the target becomes clearly visible (Fig. 6câd), and the agent proceeds with inspection. This highlights SpaceMindâs role as a decision-control hub that actively adjusts sensor parameters rather than passively consuming degraded imagery. Example 3: Prospective candidate prediction In a New Horizons search task under C2, both Standard and ReAct fail with 50-step timeouts, but Prospective succeeds in 29 steps at 2.007 m (Fig. 6eâf). The planner generates three candidate movement directions with predicted outcomes and risk scores; the selector chooses the lowest-risk option, avoiding the blind exploration that causes other modes to time out. Example 4: Self-evolution failure recovery In the New Horizons Rndz. C3 group, the baseline fails at 20 steps. After one round of self-evolution, the evolution runtime generates a learned skill encoding a distance-dependent step-size reduction strategy. The agent succeeds consistently from Round 1 onward, with step counts converging from 20 to 7 across five rounds (Fig. 6h). The learned skill is persisted as a standard skill file, making it auditable, portable, and automatically routed by the skill gateway. 18 (a) Lab: Step 1(b) Lab: Step 42(c) Exp: Step 1(d) Exp: Step 10 (e) Prosp: Step 1(f) Prosp: Step 29(g) Evol: Step 1(h) Evol: Step 8 Figure 6: Qualitative examples from the agentâs sensor view. Top row: (aâb) Lab search, where the target is not visible at Step 1 but found and approached by Step 42; (câd) inspection under C2, where the image is nearly black at Step 1 and clear after active exposure adjustment at Step 10. Bottom row: (eâf) Prospective search under C2, with an empty view at Step 1 and the target reached at Step 29; (gâh) self-evolution for New Horizons Rndz. C3, showing the baseline at Step 1 and successful approach at Step 8 after learned skill injection. 5 Discussion Framework positioning and modularity SpaceMind is designed as a modular, self-evolving embodied agent infrastructure for space operations. Its value lies in progressive layering: mod- ular knowledge, tools, and reasoning; multi-mode reasoning with hierarchical memory; skill self- evolution; and closed-loop cross-environment validation. The experimental results confirm this de- sign: the skill routing gateway achieves 100% correct task-to-skill composition across all 135 runs; the tool ablation study validates plug-and-play capability configuration; and the real-world cam- paign demonstrates zero-code-modification cross-environment transfer. Notably, the agent exhibits emergent scale awareness, autonomously reducing its step size from 1â2 m in simulation to 0.1â0.2 m in the laboratory, suggesting that the VLM can infer environmental scale from visual context. Reasoning mode selection guidelines No single reasoning mode dominates, but each has a well-defined operating regime visible only when rendezvous and search-and-approach are reported separately. For rendezvous where the target is visible, Standard achieves 13/15 or 87% with 100% success at C1 and C2; additional reasoning overhead provides no benefit. For search-and-approach where the target is not visible, Standard and ReAct both collapse to 0/5 under C2 and C3, while Prospective is the only mode that succeeds with a search subtotal of 7/15 at 47%. For inspection, ReAct achieves the highest score under C3 at 42.8 because its rapid termination limits hallucination accumulation, whereas Prospectiveâs two-stage chain compounds errors. In practice, one may default 19 to Standard and switch to Prospective upon detecting search-phase timeouts. Self-evolution mechanism insights The mechanism excels at error correction but provides lim- ited value for already-successful behavior. All four improved groups fail or score poorly at baseline, while the two successful baselines show no benefit. The mechanism is therefore most valuable as an automated failure-recovery system. Learning is fast: three groups improve significantly by Round 1 or 2 from a single failed episode. Compared with Voyager [31], Reflexion [29], and ExpeL [30], SpaceMind adds structured skill format compatibility, a quality gate with fingerprint deduplication, and operation in safety-critical physical 3D environments. Limitations and future directions Current VLMs possess limited 3D spatial understanding; SpaceMind compensates through skill knowledge and self-evolution, but the agent still occasionally misjudges spatial relationships under degraded conditions. The self-evolution mechanism provides no benefit when the baseline already succeeds, and the task-scope binding constraint prevents cross- task skill transfer, together limiting its scope to failure recovery. The laboratory validation of 18 runs demonstrates architectural feasibility but lacks statistical power for fine-grained claims, and the UE5 simulation does not incorporate orbital dynamics constraints. Future work includes training VLMs for 3D spatial reasoning in space environments, integrating orbital dynamics, and on-orbit deployment of safe agent decisions (e.g., camera exposure and sensor pointing control) where the decision-control-hub role can be validated without risking spacecraft safety. 6 Conclusions This paper presented SpaceMind, a modular and self-evolving embodied vision-language agent framework for autonomous on-orbit servicing. The framework makes four progressive contributions. First, a modular architecture decomposes agent knowledge, tools, and reasoning into independently extensible dimensions through skill modules with dynamic routing, MCP-based tools with config- urable profiles, and injectable mode skills, enabling the same codebase to operate across simulation and physical environments with zero modification. Second, three switchable reasoning modes and hi- erarchical memory provide configurable cognitive depth, with systematic evaluation across 135 runs revealing that no single mode dominates and each excels in different operating regimes. Third, a skill self-evolution mechanism enables autonomous post-deployment improvement through experience- driven knowledge generation; across 30 runs in six groups, four achieve significant positive gains, including complete failure to 100% recovery and inspection scores improving from 12 to 59, while two reveal the mechanismâs boundary conditions. Fourth, closed-loop validation across 192 runs in both UE5 simulation and a physical laboratory provides empirical support for each design decision and yields actionable engineering guidelines for reasoning mode selection and tool configuration in space agent systems. SpaceMind demonstrates that a well-structured modular agent framework, combined with au- tonomous knowledge evolution, can serve as a capable decision-control hub for complex, long-horizon space operations, thereby bridging the gap between current VLM capabilities and the demands of autonomous on-orbit servicing. Acknowledgements This work was supported by the National Key Research and Development Program of China (Grant No. 2024YFB3909300). 20 References [1] Phillip Anz-Meador. Orbital debris quarterly news. Orbital Debris Quarterly News (ODQN), 24(JSC-E-DAA-TN77633), 2020. [2] Dale Arney, Richard Sutherland, John Mulvaney, Devon Steinkoenig, Christopher Stockdale, and Mason Farley. On-orbit servicing, assembly, and manufacturing (osam) state of play, 2021. NASA OSAM state-of-play report. [3] A. Flores-Abad, O. Ma, K. Pham, and S. Ulrich. A review of space robotics technologies for on-orbit servicing. Progress in Aerospace Sciences, 68:1â26, 2014. [4] Jennifer Harbaugh and B Dunbar. On-orbit servicing, assembly, and manufacturing 2 (osam-2). STMD: Tech Demo Missions, 2022. [5] Robin Biesbroek, Sarmad Aziz, Andrew Wolahan, Stefano Cipolla, Muriel Richard-Noca, and Luc Piguet. The clearspace-1 mission: Esa and clearspace team up to remove debris. In Proc. 8th eur. conf. Sp. debris, pages 1â3, 2021. [6] Jason L Forshaw, Guglielmo S Aglietti, Nimal Navarathinam, Haval Kadhem, Thierry Salmon, AurĂŠlien Pisseloup, Eric Joffre, Thomas Chabot, Ingo Retat, Robert Axthelm, et al. Removede- bris: An in-orbit active debris removal demonstration mission. Acta Astronautica, 127:448â463, 2016. [7] Victor Rodriguez-Fernandez, Alejandro Carrasco, Jason Cheng, Eli Scharf, Peng Mun Siew, and Richard Linares. Language models are spacecraft operators. arXiv preprint arXiv:2404.00413, 2024. [8] Alejandro Carrasco, Marco Nedungadi, Victor Rodriguez-Fernandez, and Richard Linares. Vi- sual language models as operator agents in the space domain. In AIAA SCITECH 2025 Forum, page 1543, 2025. [9] Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345, 2024. [10] Xinyi Hou, Yanjie Zhao, Shenao Wang, and Haoyu Wang. Model context protocol (mcp): Landscape, security threats, and future research directions. ACM Transactions on Software Engineering and Methodology, 2025. [11] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In The eleventh international conference on learning representations, 2022. [12] Hoang Anh Dung, Bo Chen, and Tat-Jun Chin. A spacecraft dataset for detection, segmentation and parts recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2012â2019, 2021. [13] Guangyuan Zhao, Xue Wan, Yaolin Tian, Yadong Shao, and Shengyang Li. 3d component segmentation network and dataset for non-cooperative spacecraft. Aerospace, 9(5):248, 2022. 21 [14] Yadong Shao, Aodi Wu, Shengyang Li, Leizheng Shu, Xue Wan, Yuanbin Shao, and Junyan Huo. Satellite component semantic segmentation: Video dataset and real-time pyramid at- tention and decoupled attention network. IEEE Transactions on Aerospace and Electronic Systems, 59(6):7315â7333, 2023. [15] Mate Kisantal, Sumant Sharma, Tae Ha Park, Dario Izzo, Marcus Märtens, and Simone DâAmico. Satellite pose estimation challenge: Dataset, competition design, and results. IEEE Transactions on Aerospace and Electronic Systems, 56(5):4083â4098, 2020. [16] Tae Ha Park, Marcus Märtens, Gurvan Lecuyer, Dario Izzo, and Simone DâAmico. Speed+: Next-generation dataset for spacecraft pose estimation across domain gap. In 2022 IEEE aerospace conference (AERO), pages 1â15. IEEE, 2022. [17] Tae Ha Park and Simone DâAmico. Robust multi-task learning and online refinement for spacecraft pose estimation across domain gap. Advances in Space Research, 73(11):5726â5740, 2024. [18] Leo Pauly, Wassim Rharbaoui, Carl Shneider, Arunkumar Rathinam, Vincent Gaudilliere, and Djamila Aouada. A survey on deep learning-based monocular spacecraft pose estimation: Current state, limitations and prospects. Acta Astronautica, 212:339â360, 2023. [19] Matthew Foutter, Praneet Bhoj, Rohan Sinha, Amine Elhafsi, Somrita Banerjee, Christopher Agia, Justin Kruger, Tommaso Guffanti, Daniele Gammelli, Simone DâAmico, et al. Adapting a foundation model for space-based tasks. arXiv preprint arXiv:2408.05924, 2024. [20] Yujia Qin, Shengding Hu, Yankai Lin, Weize Chen, Ning Ding, Ganqu Cui, Zheni Zeng, Xu- anhe Zhou, Yufei Huang, Chaojun Xiao, et al. Tool learning with foundation models. ACM Computing Surveys, 57(4):1â40, 2024. [21] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824â24837, 2022. [22] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Ad- vances in neural information processing systems, 36:11809â11822, 2023. [23] D. Driess, F. Xia, M. S. M. Sajjadi, C. Lynch, A. Chowdhery, B. Ichter, A. Wahid, J. Tompson, Q. Vuong, T. Yu, et al. PaLM-E: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. [24] M. Ahn, A. Brohan, N. Brown, Y. Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gober, K. Gopalakrishnan, et al. Do as I can, not as I say: Grounding language in robotic affordances. In Conference on Robot Learning (CoRL), 2022. [25] W. Huang, F. Xia, T. Xiao, H. Chan, J. Liang, P. Florence, A. Zeng, J. Tompson, I. Mordatch, Y. Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. In Conference on Robot Learning (CoRL), 2023. [26] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choro- manski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision- language-action models transfer web knowledge to robotic control, 2023. URL https://arxiv. org/abs/2307.15818, 1:2, 2024. 22 [27] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. [28] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. [29] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. Advances in neural information processing systems, 36:8634â8652, 2023. [30] A. Zhao, D. Huang, Q. Xu, M. Lin, Y. Liu, and G. Huang. ExpeL: LLM agents are experiential learners. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 2024. [31] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. [32] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and service robotics: Results of the 11th international conference, pages 621â635. Springer, 2017. [33] Aodi Wu, Jianhong Zuo, Zeyuan Zhao, Xubo Luo, Ruisuo Wang, and Xue Wan. Spacesense- bench: A large-scale multi-modal benchmark for spacecraft perception and pose estimation. arXiv preprint arXiv:2603.09320, 2026. [34] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631, 2025. 23