REASIMO: Autonomy for Icy Moon Missions

• REASIMO is a mission-oriented framework for autonomous operations on icy moons like Europa and Enceladus, enabling science sampling with minimal Earth-in-the-loop support.
• It integrates AI Space Cortex modules for vision-based target selection, fault detection via MONSID, and online recalibration to maintain operational accuracy.
• Empirical tests on NASA JPL’s OWLAT testbed demonstrate REASIMO’s fault recovery capabilities and high calibration accuracy (90%-98%) in autonomous science missions.

Searching arXiv for REASIMO and closely related terms to ground the article in the cited papers. REASIMO, short for Robust, Explainable Autonomy for Scientific Icy Moon Operations, is a mission-oriented autonomy effort for future ocean-world exploration, and it is presented as the core application context for the broader AI Space Cortex architecture. It is intended to enable a landed spacecraft—especially one operating on Europa-, Enceladus-, or similarly harsh icy-moon surfaces—to carry out science sampling end-to-end with minimal or no Earth-in-the-loop intervention, while remaining robust to faults and explainable to operators (Touma et al., 9 Jul 2025).

1. Terminology and referential scope

Within the supplied arXiv material, REASIMO denotes the ocean-world autonomy effort introduced in "AI Space Cortex: An Experimental System for Future Era Space Exploration" (Touma et al., 9 Jul 2025). In that usage, the acronym expands to Robust, Explainable Autonomy for Scientific Icy Moon Operations and names a mission-oriented framework for landed icy-moon operations.

The term is, however, potentially confusable with unrelated material. In "RESOLVE: A new algorithm for aperture synthesis imaging of extended emission in radio astronomy," the query term “REASIMO” does not appear in the paper and is best interpreted as a typo or misspelling of RESOLVE, short for Radio Extended SOurces Lognormal deconVolution Estimator (Junklewitz et al., 2013). A separate paper on passive multi-RIS wireless imaging is described in the supplied synthesis as “REASIMO-style wireless regional imaging”, but the paper itself concerns wireless regional imaging through reconfigurable intelligent surfaces rather than REASIMO as an acronym (Wang et al., 2023). REASIMO should also be distinguished from REIS, short for Redundancy Elimination for Inference-efficient Systems, a dual-process framework for on-device robotic planning (Lee et al., 29 May 2026).

A common misconception, therefore, is to treat REASIMO as a generic label for imaging or reasoning systems. In the precise sense established by the supplied material, REASIMO is the NASA COLDTech-related autonomy effort for scientific icy moon operations, not the radio-imaging algorithm RESOLVE, not the passive RIS imaging framework, and not the robotic planning framework REIS.

2. Mission setting and operational rationale

REASIMO is framed by the operational realities of ocean worlds such as Europa and Enceladus. These missions are characterized by long communication delays, limited power, radiation and hostile environmental conditions, and a short mission lifetime. For Europa-class missions, the paper notes that direct-Earth communication from the surface could be available only for a limited fraction of each orbit, with large latency. Under these conditions, onboard autonomy is not optional (Touma et al., 9 Jul 2025).

The operational requirement is broader than nominal sequencing. A spacecraft must choose where to sample, verify that a site is mechanically safe and scientifically promising, recover from anomalies, and continue operating even when the ground cannot intervene in time. The paper explicitly contrasts this with the traditional spacecraft paradigm of hard-coded, deterministic task logic and safe-mode-centric fault handling. Older spacecraft autonomy is described as scripted, state-machine based, conservative, and oriented toward spacecraft survival rather than continued science.

REASIMO instead aims for pre-trained behaviors rather than hard-coded predetermined logic, mission reasoning based on telemetry + vision + LLMs, AI-assisted decision-making, and autonomous recovery strategies that attempt to keep the mission moving. The operational meaning of “without Earth-in-the-loop” is not absolute independence from Earth in all cases; rather, the spacecraft should be able to detect faults, understand what they mean, choose a recovery path, continue mission operations, and only escalate to Earth when necessary. The mission philosophy is summarized in the supplied material as keeping the spacecraft scientifically productive autonomously, rather than merely safe (Touma et al., 9 Jul 2025).

3. AI Space Cortex architecture and control philosophy

REASIMO contributes to NASA’s COLDTech program and is implemented as the AI Space Cortex, described as an “intelligent framework for autonomous space exploration.” The architecture comprises five major functional parts (Touma et al., 9 Jul 2025).

Hierarchical Controller (HC): This module executes task-level robotic operations, handles nominal task sequencing, and translates AI decisions into concrete arm motions, probing, scooping, and storage operations. It is ROS Noetic-based, uses ROS topics, services, and action servers, handles state transitions, executes mission sequencing, monitors joint health, power, and sensor status, and can re-plan if the selected sampling site proves unsuitable.

Intelligent Scene Interaction (ISI): This subsystem performs vision-based perception and semantic understanding. It uses RGB-D imagery, segmentation, and LLM reasoning to select scientific targets. Its pipeline acquires an RGB-D image, applies SAM-1 segmentation, filters objects by size, proximity, material class, and reachability, and sends refined candidates to GPT-4o for scientific ranking and justification.

Explanation Engine (EE): This is the human-machine interface. It visualizes segmentations, overlays confidence colors—green = high confidence/high viability, orange = medium confidence, red = non-viable—streams telemetry, exposes live fault indicators, supports operator override, and reformats LLM text into structured rationale.

MONSID fault detection system: This is Model-based Off-Nominal State Identification and Detection. It detects, isolates, and reports faults in the arm and sensors through a model-based analytic redundancy scheme.

Online recalibration module: This module autonomously recalibrates kinematics after fault-induced misalignment and serves as a recovery capability for a damaged or miscalibrated manipulator.

A distinctive aspect of the framework is that it is personality-driven. The autonomy style can be influenced with modes such as Conservative Mode, Scientific Curiosity Mode, and Adventurous Mode. The supplied material describes this as a promptable behavioral style; operators can say, for example, “be more aggressive for the next 72 hours,” and the system adapts its exploration strategy accordingly. This is presented as a departure from fixed scripting, because mission-level behavioral adaptation is achieved without rewriting onboard rules (Touma et al., 9 Jul 2025).

4. Autonomous science sampling workflow

The paper presents an end-to-end autonomous science mission sequence organized around verification, selection, probing, collection, and logging (Touma et al., 9 Jul 2025).

Pre-mission checks: The system verifies battery health, thermal status, sensor readiness, fault-monitor readiness, and environmental feasibility.

Scene analysis and target selection: The mast camera captures RGB-D images. These are processed by SAM-1 segmentation, then filtered geometrically and semantically, and then passed to an LLM for scientific prioritization. The filtering stage includes some hard-coded geometric constraints based on end-effector geometry, which remove impossible targets and reduce compute burden.

Probing and material analysis: A cone penetrometer is used to probe candidate sites and measure force/penetration behavior. This functions as a physics-based validation step: a site that appears scientifically promising according to the LLM must still be validated for scoopability and safety.

Sample collection and storage: The arm swaps from penetrometer to scoop, performs a scoop maneuver, and deposits material in a cache.

Mission completion and telemetry logging: The system records site-selection rationale, probing data, cached sample properties, and mission logs, which are transmitted or exposed through the explanation interface.

The workflow is explicitly multimodal. It combines semantic scene understanding and LLM-based site ranking with force-torque probing and HC-based robotic execution. A plausible implication is that REASIMO is designed not as a purely language-conditioned planner, but as a hybrid autonomy stack in which perceptual-semantic prioritization is constrained by geometric feasibility and mechanical validation.

5. Fault diagnosis, local recovery, and online recalibration

REASIMO’s resilience is organized around MONSID, the HC’s escalating recovery logic, and a non-parametric, data-efficient, online recalibration method on pose space (Touma et al., 9 Jul 2025).

MONSID is model-based and uses analytic redundancy: the system propagates measurements through both forward and reverse paths in a model, then checks consistency between multiple estimates of the same physical quantity. If inconsistencies persist beyond tuned thresholds, a fault is declared. In the OWLAT-specific model, MONSID monitors joint actuator commands, encoder positions, joint velocities, the kinematic model, and end-effector pose from camera view. It uses ambiguity groups, and the supplied material states that the OWLAT topology allows distinction between actuator faults and encoder faults for each joint pair, plus a separate ambiguity group involving the kinematics component and end-effector pose sensor.

Recovery is stratified. Minor faults trigger retry or restart; medium faults pause execution and request AI Space Cortex intervention; critical but fixable faults halt motion and trigger recovery such as recalibration; severe/unfixable faults enter safe mode. The paper is explicit that safe mode remains a fallback, but is insufficient by itself for time-limited ocean-world missions because it protects hardware while often stopping science.

The recalibration module operates on $\mathbb{S}^3 \times \mathbb{R}^3$, meaning orientation and position are handled directly in pose space rather than only in joint space. The quaternion geodesic distance is given by

$$d_{\mathbb{S}^3}(\mathbf{q}_1, \mathbf{q}_2) = 2 \cos^{-1} \left( \left| \langle \mathbf{q}_1, \mathbf{q}_2 \rangle \right| \right).$$

The GP posterior mean and variance are

$$\mu_*(x^*) = \mu(x^*) + K(x^*,X)\tilde K(X,X)^{-1}(Y-\mu(X)),$$

$$\sigma_*^2(x^*) = k(x^*,x^*) - K(x^*,X)\tilde K(X,X)^{-1}K(X,x^*),$$

with

$\tilde K = K + \sigma_\epsilon^2 I.$

A valid product kernel on $\mathbb{S}^3 \times \mathbb{R}^3$ is constructed as

$$k_{\mathbb{S}^3\times\mathbb{R}^3}(x_i,x_j)=\sigma_s^2\,k_{\mathbb{S}^3}(q_i,q_j)\,k_{SE}(p_i,p_j).$$

The calibration objective combines position and orientation error:

$$f(p,q)=- \left( \alpha_1 \frac{f_p(p)}{\sup |f_p|} + \alpha_2 \frac{f_q(q)}{\sup |f_q|} \right),$$

where $f_p(p)=\|p-\tilde p\|$, $f_q(q)=d_{\mathbb{S}^3}(q,\tilde q)$, and $$d_{\mathbb{S}^3}(\mathbf{q}_1, \mathbf{q}_2) = 2 \cos^{-1} \left( \left| \langle \mathbf{q}_1, \mathbf{q}_2 \rangle \right| \right).$$0.

The next calibration pose is selected by GP-UCB:

$$d_{\mathbb{S}^3}(\mathbf{q}_1, \mathbf{q}_2) = 2 \cos^{-1} \left( \left| \langle \mathbf{q}_1, \mathbf{q}_2 \rangle \right| \right).$$1

After measurement, the DH parameter correction is estimated with a quadratic program:

$$d_{\mathbb{S}^3}(\mathbf{q}_1, \mathbf{q}_2) = 2 \cos^{-1} \left( \left| \langle \mathbf{q}_1, \mathbf{q}_2 \rangle \right| \right).$$2

Algorithmically, the loop repeatedly chooses the next pose by GP-UCB, moves the arm, measures joint variables and end-effector pose, computes position and orientation errors, updates the GP, and refines DH parameters via QP. In the supplied description, this is the online calibration loop REASIMO uses for recovery.

6. Testbed, empirical results, limitations, and implications

The experiments are performed at NASA JPL on the OWLAT testbed, the Ocean Worlds Lander Autonomy Testbed (Touma et al., 9 Jul 2025). OWLAT includes a 7-DoF WAM robotic arm, a 6-DoF Stewart platform to emulate lander base dynamics, a 2-DoF vision system / mast camera, force-torque sensing, geotechnical tools including a cone penetrometer, shear bevameter, and pressure sinkage plate, sampling tools including scoops and a swarf-collecting drill, and a 1 m² simulant workspace with terrain diversity. The workspace includes WF-34 quartz sand and MPACS. A simulator using DARTS supports higher-risk development before hardware deployment. The AI Space Cortex runs on an off-the-shelf compact system comprising an Intel NUC with 8-core CPU, 64 GB RAM available, though 8 GB was sufficient, an NVIDIA RTX 2060 / 6 GB GPU, and an Intel RealSense D415 camera.

Two major experiments are reported. In Test 1, a 0.5235 radian bias in joint 7 is manually injected. MONSID detects the encoder/kinematic inconsistency and the system halts motion. The recalibration module then runs across 38 distinct arm configurations; the final corrected joint offset is about 0.57 rad; the reported accuracy measure is

$$d_{\mathbb{S}^3}(\mathbf{q}_1, \mathbf{q}_2) = 2 \cos^{-1} \left( \left| \langle \mathbf{q}_1, \mathbf{q}_2 \rangle \right| \right).$$3

The calibration phase lasts 31 minutes, and the entire mission including detection, recovery, and sample delivery takes 39 minutes. The paper reports that the arm returned to operational status and successfully completed the scoop mission after recalibration. It also notes around 10 trials, calibration accuracy between 90% and 98%, and convergence in 15 to 40 sampling points.

In Test 2, the full autonomy stack is evaluated. A single RGB-D frame is segmented using SAM-1; the system evaluates 7 candidate sites; each site receives a confidence score from 1–10; three are non-viable, two are medium confidence, one is high-priority. The HC executes force-torque probing, with reported penetration depths including 0.067 m at the best site, along with 0.0014 m and 0.0017 m at other locations. The probing phase takes 230 seconds; scoop and cache take about 124 seconds; the full mission completes in about 8 minutes and 40 seconds, with segmentation under 3 seconds and LLM evaluation at about 2.5 seconds per site.

The paper also states several limitations. GPT-4o API usage is not flight-realistic by itself. Future spaceflight would require smaller, possibly local models and stronger optimization or quantization. Cloud/API dependence is not directly space-deployable. LLM outputs are probabilistic and difficult to quantify statistically compared to classical models. Hard-coded heuristics still exist in filtering and some mission constraints. The system is validated on a flight-relevant testbed, not on an actual Europa or Enceladus mission. Safe mode remains a fallback, and full autonomy does not eliminate the need for human oversight in all cases.

The broader implication stated in the supplied material is that future ocean-world missions may need an autonomy stack that is strongly autonomous, fault-tolerant, scientifically aware, explainable, and able to operate for long periods without ground control. In that sense, REASIMO is presented not merely as a software subsystem, but as a concrete experimental demonstration that AI-assisted, explainable, onboard autonomy can support the operational realities of ocean-world exploration (Touma et al., 9 Jul 2025).