A Computational Framework for Cross-Domain Mission Design and Onboard Cognitive Decision Support

This paper introduces a unified computational framework featuring the Autonomy Necessity Score to quantify decision constraints across seven diverse mission architectures and validates the viability of onboard LLM-based cognitive decision support, demonstrating that foundation models can achieve 80% accuracy within strict latency budgets for high-autonomy space and underwater operations.

The Gist

Imagine you are the captain of a ship, but your ship is so far away that by the time you shout a command to your crew, the message takes hours to arrive, and their reply takes hours to get back to you. In space, this is the reality. The speed of light is fast, but space is really big. If you are on Mars or near Saturn, you can't just radio Earth for help when something goes wrong; you have to make the decision yourself, instantly.

This paper is like a universal "Independence Test" for space missions. It asks a simple question: "How much does this robot need to think for itself before its human boss can even say 'hello'?"

Here is the breakdown of their work, explained simply:

1. The "Autonomy Necessity Score" (The Independence Meter)

The authors invented a new score called the Autonomy Necessity Score (ANS). Think of this like a "maturity meter" for a robot.

• Low Score (0.0): The robot is like a child holding a parent's hand. It can call home for every little decision (like a satellite orbiting Earth).
• High Score (0.8+): The robot is like an explorer alone in the wilderness. It must be a grown-up because the parent is too far away to help (like a probe near Saturn).

They tested this meter on seven different types of missions, ranging from satellites watching Earth to underwater robots clearing mines, and finally to buoys floating on the methane lakes of Titan (a moon of Saturn).

2. The Seven "Characters" in the Story

To prove their meter works, they looked at seven very different scenarios:

• The Watchdogs (SCOPE & H.S.A.D.S.): Satellites watching Earth for fast-moving missiles. They are close to home, so they can talk to Earth quickly.
• The Underwater Swimmers (AHMS): A team of five underwater robots working together to clear mines. They can't use GPS underwater, so they have to be very smart and coordinated.
• The Martian Navigators (MarsNav & EDL): Satellites helping landers on Mars. When a lander is crashing down to the surface, it has to make decisions in minutes. Earth is too far away to help.
• The Deep Space Swarm (ChipSat): A hundred tiny satellites near Jupiter. They are so far away that a message takes hours. They must solve problems instantly.
• The Titan Buoys: Floating sensors on a moon of Saturn. They are the most isolated of all.

3. The "Aha!" Moments (What They Discovered)

By running complex math simulations on all these missions at once, they found three surprising rules that you wouldn't see if you only looked at one mission:

• The "Fat Battery" Rule: For the underwater robots, they realized that to carry enough battery power, the robot's body must be at least 1 meter wide. If it's smaller, it literally can't hold the battery. It's like trying to fit a suitcase in a backpack that's too small; the math says "no."
• The "Silent Radio" Rule: When a Mars mission happens during a "solar conjunction" (when the Sun is between Earth and Mars), the radio signal gets scrambled by solar noise. The team realized the robot must automatically switch to a slower, safer speed without waiting for permission from Earth, or it will lose contact forever.
• The "Perfect Timing" Rule: The underwater robots need to coordinate their movements with extreme precision. The team found that the robots need to sync their clocks 2.4 times better than originally planned, or they won't be able to mimic the sound of a ship to trick the mines.

4. The "AI Brain" Experiment

The most exciting part is the second half of the paper. They asked: "Can we put a super-smart AI (like a very advanced chatbot) inside the spaceship to help make these hard decisions?"

They took three of the world's most powerful AI models (Llama, DeepSeek, and Qwen) and fed them 10 different "What would you do?" scenarios based on their math.

• The Test: They simulated a crisis (like a clock breaking or a power drop) and asked the AI to choose the best action.
• The Result: The best AI got 80% of the answers right.
• The Catch: The AI was sometimes too optimistic (thinking it could keep working when it should have stopped) or didn't understand the specific math trade-offs. However, it was fast enough (under 2 seconds) to actually be used on a real spaceship.

The Big Takeaway

This paper is a blueprint for the future of space exploration. It tells us that as we go further out into the solar system, our robots can't just be remote-controlled toys; they need to be independent thinkers.

The authors created a scorecard to tell engineers exactly how independent a robot needs to be. Furthermore, they proved that AI is ready to be the "co-pilot" for these deep-space missions, helping the robots make life-or-death decisions when the humans are too far away to help.

In short: Space is too big for us to hold the robot's hand. We need to build robots that are smart enough to walk alone, and we now have a map and a test to make sure they are ready for the journey.

Technical Summary

1. Problem Statement

The deployment of distributed autonomous systems in deep space and remote environments (e.g., Mars, Titan, underwater) faces a fundamental engineering constraint: communication latency. As the round-trip time (RTT) between an asset and ground operators increases, the ability to delegate critical decisions to the ground diminishes.

• The Gap: Current literature typically analyzes missions in isolation (e.g., optimizing orbital mechanics for Earth or reliability for a specific Mars lander). There is no unified framework to quantify how much autonomy a mission requires or to compare the autonomy demands of heterogeneous systems (underwater swarms vs. deep-space constellations) on a common scale.
• The Risk: Over-engineering ground links for high-latency missions wastes mass and power, while under-designing onboard decision-making for time-critical events risks mission failure.

2. Methodology

The authors propose a unified computational framework validated through a multi-stage approach:

A. The Autonomy Necessity Score (ANS)

The core methodological contribution is the Autonomy Necessity Score (ANS), a dimensionless, log-domain metric defined as:
$$ANS(\tau) = \min\left(1, \frac{\log_{10}(\tau + 1)}{5}\right)$$
Where $\tau$ is the round-trip communication latency in seconds.

• Scale: Maps systems from 0.0 (fully ground-supervised) to 1.0 (fully autonomous).
• Calibration: The denominator (5) ensures the score approaches unity only beyond $10^5$ seconds, preventing saturation within the studied mission set.
• Application: The ANS is calculated per mission phase; the worst-case latency determines the minimum required autonomy architecture.

B. Seven Heterogeneous Mission Architectures

The framework is applied to seven distinct mission profiles spanning five engineering domains:

1. SCOPE: GEO+LEO infrared surveillance constellation (Hypersonic cueing).
2. H.S.A.D.S.: LEO dual-band tracking satellite with onboard Extended Kalman Filter (EKF).
3. AHMS: Autonomous Heavy Minesweeping System (5-vehicle underwater swarm).
4. MarsNav: 6-satellite Walker-Delta Mars navigation constellation.
5. MOC/EDL: Mars Entry, Descent, and Landing operations.
6. ChipSat: 100-node deep-space inter-satellite link (ISL) swarm near Jupiter.
7. Titan: 4-buoy drifting network for Kraken Mare exploration.

C. Nine Computational Studies

To validate the ANS and derive physical constraints, the authors executed nine independent computational modules across all missions:

1. Walker spherical-cap coverage mechanics.
2. Infrared detection (Neyman–Pearson optimality).
3. EKF hypersonic tracking state estimation.
4. TDMA Monte Carlo science-yield sensitivity.
5. Cross-mission RF and acoustic link budget analysis.
6. Latency–autonomy mapping (ANS calculation).
7. Power budget sizing (Solar, RTG, LFP batteries).
8. Distributed magnetic-signature formation emulation.
9. Arrhenius-corrected cryogenic swarm reliability.

D. LLM-Based Decision Support Evaluation

Building on the physics-grounded data, the authors evaluated an Autonomous Mission Decision Support (AMDS) layer.

• Models Tested: Llama-3.3-70B, DeepSeek-V3, and Qwen3-A22B.
• Setup: 10 structured anomaly scenarios derived directly from the computational studies were queried via the Nebius AI Studio API.
• Metrics: Decision accuracy against physics-grounded truth, confidence calibration, and inference latency.

3. Key Contributions

1. Unified Metric (ANS): Introduction of a continuous, log-domain metric that quantifies the minimum autonomy forced by physics (latency) rather than operational preference.
2. Cross-Domain Validation: A comprehensive validation suite covering seven diverse architectures and nine engineering disciplines, revealing constraints invisible to single-mission analysis.
3. Onboard AI Feasibility: Demonstration that Large Language Models (LLMs) can function as an onboard cognitive layer for high-ANS missions, achieving high accuracy within strict latency budgets.

4. Key Results

A. Physical and Design Constraints

The cross-mission analysis yielded three "physically non-obvious" constraints:

• AHMS Battery Mass: A 2,431 kg battery mass is required for the underwater swarm, constraining the vehicle hull diameter to $\ge$ 1.0 m. A 0.5 m hull is physically infeasible due to power density limits.
• Mars Solar Conjunction: The Mars HGA-to-Earth link fails at solar conjunction (−1.5 dB margin). The only recovery is an autonomous rate reduction (64 kbps $\to$ 4 kbps), which must be pre-committed without ground intervention.
• AHMS Timing Synchronization: To maintain coherent acoustic signature emulation, the timing accuracy requirement must be tightened from the original 100 ms to 41 ms (a 2.4$\times$ increase), invalidating the original subsystem specification.

B. ANS Spectrum Analysis

The missions cluster into three distinct autonomy regimes:

• Low Latency (ANS $\approx$ 0.06): SCOPE and H.S.A.D.S. (Earth LEO). Ground supervision is possible, but onboard complexity is driven by real-time signal processing.
• Medium Latency (ANS $\approx$ 0.58): MarsNav and MOC/EDL. Autonomy is mandated by the 7-minute EDL window.
• High Latency (ANS > 0.74): ChipSat (Jupiter) and Titan. Autonomy is dictated entirely by distance (up to ~3 hours RTT).

C. LLM Decision Support Performance

• Accuracy: Llama-3.3-70B achieved 80% decision accuracy, significantly outperforming a conservative heuristic baseline (0%).
• Latency: All 180 inference calls completed in a mean of 1.96 seconds, fitting within the 2-second budget for radiation-hardened edge deployment.
• Failure Modes:
  • Quantitative Trade-offs: Models failed on scenarios requiring strict numerical trade-offs (e.g., ChipSat clock drift) where the "correct" physics answer (tolerate drift) was less intuitive than the "corrective" answer.
  • Optimism Bias: DeepSeek showed a bias toward continuing full science operations despite power deficits.
• Calibration: Confidence scores were poorly calibrated (near-flat), suggesting that ensemble majority voting (at $T=0.05$) is a safer deployment strategy than relying on single-model confidence thresholds.

5. Significance

• Methodological Shift: Moves aerospace engineering from isolated mission design to a cross-domain analytical framework, allowing engineers to systematically allocate resources (power, compute, comms) based on the actual autonomy gradient of a mission.
• AI in Space: Provides empirical evidence that foundation models can serve as viable onboard cognitive layers for deep-space missions, provided they are integrated with physics-grounded constraints and ensemble methods.
• Design Implications: Highlights that autonomy requirements are not just about "smart" software but are often dictated by hard physical limits (battery mass, link budgets, acoustic coherence) that can only be identified through unified system analysis.

This work establishes a foundational methodology for the next generation of autonomous systems, bridging the gap between traditional systems engineering and emerging AI-driven decision support.