Beyond Digital Knowledge: How Artificial Intelligence Could Learn Through Physical Experience

Introduction

The remarkable progress of artificial intelligence has been built upon a single, fundamental assumption: knowledge is acquired from digital information. Large language models have learned from books, scientific papers, websites, software repositories, images, audio, and videos. Yet this represents only one dimension of intelligence.

Human intelligence develops through a continuous interaction with the physical world. We learn not only from language but also from touching objects, observing natural phenomena, manipulating tools, perceiving gravity, temperature, pressure, texture, light, sound, and the countless subtleties of social interaction. This raises a fascinating question:

Could future AI systems acquire knowledge by directly experiencing the physical world rather than merely reading about it?

The answer may define the next major stage in artificial intelligence.

---

The Limits of Purely Digital Learning

Current AI systems possess extraordinary reasoning abilities but remain fundamentally indirect observers of reality.

They know what fire is because millions of documents describe it.

They understand gravity because physics textbooks explain it.

They recognize emotions because countless conversations contain emotional language.

However, they have never directly sensed:

• the heat emitted by a flame,

• the acceleration of a falling object,

• the vibration of a machine,

• the smell of smoke,

• the resistance of a material,

• or the uncertainty inherent in real-world sensory measurements.

Their understanding is statistical rather than experiential.

In contrast, biological intelligence integrates language with direct sensory experience throughout life.

---

Toward Embodied Artificial Intelligence

The next frontier may be Embodied AI—artificial intelligence connected to the physical world through sensors and actuators.

Instead of existing solely inside data centers, AI could receive continuous streams of real-world information from instruments such as:

• High-resolution cameras

• Microphones

• Thermal sensors

• Spectrometers

• Accelerometers

• Pressure sensors

• Chemical detectors

• Weather stations

• Microscopes

• Telescopes

• Medical imaging devices

• Industrial sensors

• Robotic manipulators

These devices would become the AI's equivalent of sensory organs.

Rather than reading a description of rainfall, the AI could simultaneously observe:

• atmospheric pressure,

• humidity,

• cloud formation,

• radar images,

• lightning activity,

• wind velocity,

• and ground-level precipitation.

Its knowledge would emerge from direct measurement instead of textual description.

---

Learning Through Scientific Instruments

Perhaps the greatest opportunity lies not in robotics but in scientific instrumentation.

Imagine an AI permanently connected to laboratories around the world.

It could observe in real time:

• chemical reactions,

• crystal growth,

• plasma formation,

• protein folding,

• astronomical events,

• volcanic activity,

• ocean currents,

• particle collisions,

• biological cell division.

Unlike human scientists, whose observations are necessarily intermittent, AI could monitor every variable continuously for years.

Patterns invisible to human researchers might become apparent simply because no observation would ever be interrupted.

Such systems could accelerate scientific discovery across nearly every discipline.

---

Understanding Human Behavior Through Continuous Observation

Language represents only a fraction of human communication.

Psychologists estimate that much interpersonal communication occurs through nonverbal cues, including facial expressions, posture, gestures, tone of voice, and interpersonal distance.

An AI capable of observing these signals directly—under ethical safeguards and with informed consent—could develop a far richer understanding of human behavior than one trained exclusively on text.

Future systems might learn:

• how trust develops,

• how cooperation emerges,

• how conflicts escalate,

• how leadership forms,

• how children learn,

• how experts solve problems,

• and how cultural norms evolve.

This would represent experiential social learning rather than textual analysis.

---

Building a Digital Nervous System

To achieve such capabilities, AI would require an architecture resembling a biological nervous system.

Instead of processing isolated prompts, it would continuously integrate multiple sensory streams.

A possible architecture might include:

Perception Layer

Collects information from cameras, microphones, environmental sensors, satellites, robots, microscopes, and industrial equipment.

Sensor Fusion Layer

Combines observations from multiple sources into coherent representations of physical reality.

World Model

Constructs an evolving simulation of the surrounding environment, continuously updated by incoming observations.

Reasoning Layer

Generates hypotheses explaining observed phenomena.

Experimentation Layer

Designs new observations or experiments to test competing hypotheses.

Learning Layer

Updates internal models according to empirical evidence rather than solely statistical correlations.

This resembles the scientific method more than conventional machine learning.

---

Active Learning Through Experimentation

Humans rarely learn passively.

Children drop objects repeatedly to understand gravity.

Scientists manipulate variables to discover causal relationships.

Future AI systems may similarly become active learners.

Rather than waiting for new data, they could:

• adjust microscope settings,

• reposition robotic cameras,

• modify laboratory conditions,

• perform repeated measurements,

• design new experiments,

• identify anomalous observations,

• refine hypotheses through iterative testing.

This transition—from passive prediction to active experimentation—may represent one of the most significant advances in AI.

---

From Correlation to Causality

Today's AI excels at identifying statistical relationships.

Future embodied AI could move closer to understanding causality.

By directly manipulating variables and observing outcomes, AI might distinguish:

• coincidence from causation,

• stable laws from temporary correlations,

• genuine physical mechanisms from observational artifacts.

Such capability would transform AI from an information processor into a scientific investigator.

---

Challenges and Ethical Considerations

This vision also introduces profound challenges.

Continuous observation raises legitimate concerns regarding:

• privacy,

• informed consent,

• surveillance,

• data ownership,

• cybersecurity,

• manipulation,

• algorithmic bias.

Any system capable of observing human life at large scale must operate within transparent legal and ethical frameworks.

Equally important is ensuring that AI-generated scientific conclusions remain reproducible and independently verifiable.

---

Toward Experiential Artificial Intelligence

If today's large language models represent digital intelligence, future systems may evolve into experiential intelligence.

Their knowledge would emerge from the combination of:

• digital information,

• physical observation,

• autonomous experimentation,

• multisensory perception,

• continuous interaction with reality.

Such systems would no longer rely solely on humanity's written descriptions of the world.

Instead, they would become direct observers of nature.

---

Conclusion

Artificial intelligence has reached extraordinary levels of capability by learning from humanity's accumulated digital knowledge. Yet this represents only the beginning of what machine intelligence may become.

The next great leap may occur when AI moves beyond reading about reality and begins to experience it through networks of scientific instruments, robotic systems, environmental sensors, and ethically governed observations of human activity.

This evolution would not imply consciousness in the human sense. Rather, it would represent a profound expansion in how artificial systems acquire knowledge—shifting from second-hand descriptions to first-hand empirical evidence.

Just as the invention of the telescope transformed astronomy and the microscope revolutionized biology, the integration of AI with the physical world could redefine the nature of scientific discovery itself.

The future of artificial intelligence may therefore depend not only on larger models or more powerful processors, but on something far more fundamental: the ability to observe, interact with, and learn directly from reality.

 

Glossary

Active Learning
A machine learning paradigm in which an AI system actively selects the data or experiments that are expected to provide the greatest improvement in its knowledge.

Artificial General Intelligence (AGI)
A theoretical form of artificial intelligence capable of performing intellectual tasks across diverse domains at a level comparable to or exceeding that of humans.

Causal Reasoning
The ability to infer cause-and-effect relationships rather than merely identifying statistical correlations between variables.

Embodied AI
A field of artificial intelligence in which an intelligent system interacts with the physical world through sensors and actuators, enabling learning from real-world experience.

Empirical Learning
Knowledge acquisition based on direct observation, measurement, experimentation, and interaction with the environment.

Experiential Artificial Intelligence (EAI)
A proposed paradigm in which AI systems acquire knowledge through continuous interaction with the physical world using scientific instruments, sensors, robotic platforms, and autonomous experimentation. Unlike traditional AI, which primarily learns from digital data, EAI integrates first-hand sensory observations into its learning process.

Foundation Model
A large-scale AI model trained on extensive datasets that can be adapted to perform a wide range of downstream tasks.

Grounded Intelligence
An approach to AI in which concepts are linked to real sensory experiences rather than existing solely as symbolic or linguistic representations.

Multimodal Learning
Learning from multiple forms of information simultaneously, such as text, images, audio, video, sensor measurements, and environmental data.

Perception
The process of acquiring and interpreting information from sensors in order to understand the surrounding environment.

Physical World Model
An internal computational representation that allows an AI system to predict, explain, and reason about physical phenomena based on sensory observations.

Robotic Manipulation
The capability of robotic systems to interact physically with objects in their environment through controlled movements.

Scientific Instrumentation
Specialized devices used to measure, observe, and record physical, chemical, biological, or astronomical phenomena.

Sensor Fusion
The integration of data from multiple sensors into a unified representation that provides greater accuracy and reliability than any individual sensor alone.

Self-Supervised Learning
A learning approach in which models generate supervisory signals from the structure of the input data itself, reducing dependence on manually labeled datasets.

World Model
An internal representation of the external environment that enables prediction, planning, simulation, and reasoning about future events.

---

References

Barsalou, L. W. (1999). Perceptual Symbol Systems. Behavioral and Brain Sciences, 22(4), 577–660.

Brooks, R. A. (1991). Intelligence Without Representation. Artificial Intelligence, 47(1–3), 139–159.

Brooks, R. A. (1986). A Robust Layered Control System for a Mobile Robot. IEEE Journal of Robotics and Automation, 2(1), 14–23.

Friston, K. (2010). The Free-Energy Principle: A Unified Brain Theory? Nature Reviews Neuroscience, 11(2), 127–138.

Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press.

Ha, D., & Schmidhuber, J. (2018). World Models. arXiv:1803.10122.

Kaelbling, L. P., & Lozano-Pérez, T. (2011). Hierarchical Task and Motion Planning in the Now. Proceedings of the IEEE International Conference on Robotics and Automation.

Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building Machines That Learn and Think Like People. Behavioral and Brain Sciences, 40, e253.

LeCun, Y. (2022). A Path Toward Autonomous Machine Intelligence. Open Review.

Marr, D. (1982). Vision: A Computational Investigation into the Human Representation and Processing of Visual Information. W. H. Freeman.

Pearl, J. (2009). Causality: Models, Reasoning, and Inference (2nd ed.). Cambridge University Press.

Russell, S., & Norvig, P. (2021). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.

Silver, D., Sutton, R. S., & Müller, M. (2024). Welcome to the Era of Experience. Nature Machine Intelligence.

Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction (2nd ed.). MIT Press.

Tenenbaum, J. B., Kemp, C., Griffiths, T. L., & Goodman, N. D. (2011). How to Grow a Mind: Statistics, Structure, and Abstraction. Science, 331(6022), 1279–1285.

Thrun, S., Burgard, W., & Fox, D. (2005). Probabilistic Robotics. MIT Press.

---

Suggested Further Reading

• Bengio, Y. (2024). Research on World Models, Agentic AI, and Autonomous Learning.
• LeCun, Y. Research on Joint Embedding Predictive Architectures (JEPA).
• Pearl, J. Research on causal inference and scientific reasoning.
• OpenAI. Research on multimodal foundation models and embodied AI.
• Google DeepMind. Research on robotics, world models, and autonomous agents.
• NVIDIA Research. AI for robotics, digital twins, and physical simulation.
• MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). Research on embodied intelligence and human-robot interaction.

 

On Imminent: Inside the Pentagon's Hunt for UFOs, by Luis Elizondo

 

The Pentagon, Faith, and the Silence

On Imminent: Inside the Pentagon's Hunt for UFOs, by Luis Elizondo

Luis Elizondo wants us to believe him for the same reason much of the American press decided, for years, that it should: because he says he was inside. "Imminent: Inside the Pentagon's Hunt for UFOs" presents itself as the testimony of a man who walked away from a twenty-year career in military intelligence on a matter of conscience — the conviction that the United States government is withholding from the public evidence that we are not alone in the universe — and the book depends, on every page, on the reader accepting that premise of authority before examining anything else. It is a risky bet for a memoir to make, and it is also, I'm afraid, the bet that ultimately gives it away. 

Elizondo, a former Army counterintelligence agent with tours in Afghanistan and Guantánamo Bay, claims to have run the Advanced Aerospace Threat Identification Program (AATIP), the Pentagon initiative created in 2007 at the urging of then-Senator Harry Reid to investigate unidentified aerial phenomena. He resigned in 2017 with a letter to then-Defense Secretary Jim Mattis denouncing "excessive secrecy" and the "institutional ridicule" directed at those who studied the subject. That resignation, followed by the leak to The New York Times of the now-famous "Tic Tac" and "Go Fast" videos, turned him overnight into the public face of the movement for disclosure on unidentified anomalous phenomena, or UAP. The trouble — and here it pays to be precise, since the book itself is not — is that the Pentagon has never confirmed this version of events. In 2019, a spokesperson told The Intercept that Elizondo "had no responsibilities with regard to the AATIP program" during his time in the Office of the Under Secretary of Defense for Intelligence. In 2021, another spokesperson repeated the same denial. Internal emails declassified years later show former superiors echoing, almost word for word, the same formula of disavowal. Elizondo disputes all of it; Harry Reid, before his death, backed his account. None of this dispute appears anywhere in "Imminent," and that silence is itself a piece of information about the book in front of you.

What the book does offer is a narrative carefully built across two registers that rarely sit comfortably together. The first is that of the disciplined intelligence officer: short sentences, precise chronologies, a bureaucratic vocabulary that evokes classified briefings more than literary prose. Elizondo is genuinely effective at describing the mechanisms by which an institution like the Department of Defense buries an uncomfortable subject without any need for centralized conspiracy: personnel reassignments, budgets diluted across adjacent programs, the informal discrediting of whoever pushes too hard. Anyone who has ever worked inside a large bureaucracy — military, corporate, or academic — will recognize the pattern, and it is in these passages, devoted to Pentagon office politics, that the book reaches its greatest plausibility. Elizondo does not need to convince us that interstellar craft exist in order to convince us that an institution might prefer silence to discomfort; that argument stands on its own, and it is a shame the author does not trust it enough.

Because the book's second register is something else entirely. It is the register of revelation, of the curtain pulled back to show what lay behind it, and there the prose grows less disciplined and more anxious to persuade. Elizondo recounts his time in AATIP's predecessor program, AAWSAP, funded by contractor Robert Bigelow and tied to the notorious Skinwalker Ranch in Utah, a property that UFO literature has surrounded for decades with paranormal anecdote — lights, cattle mutilations, apparitions. The book does not simply mention this episode as a biographical curiosity; it presents it as evidence. There are pages devoted to "remote viewing," the alleged ability to perceive distant targets with the mind, offered not as the vestige of a Cold War program the intelligence community itself discredited long ago, but as a serious clue to the nature of the UAP phenomenon. This is the point at which a book that had, until then, asked to be read as the testimony of a serious official instead begins to demand the kind of faith that esoteric literature requires. Elizondo would like both registers to coexist without friction — the disciplined analyst and the institutional seer — but an attentive reader cannot help wondering why the first, with its insistence on verifiable evidence, never applied the same rigor to the second.

Nor does the book offer the one thing that might settle the dispute over its own credibility: proof. Elizondo repeats, with an almost liturgical consistency, that the sensor data supporting his claims remains classified, and that the skepticism of independent analysts — most often Mick West, a researcher who has methodically dismantled several of the phenomenon's most widely circulated images — stems from their lack of access to that information. It is an argument that cannot be refuted because it cannot be verified either, and that is precisely its rhetorical function: it converts the absence of public evidence into further proof of the scale of the secret, a maneuver any reader familiar with conspiratorial rhetoric will recognize instantly, even coming here from a source with far more respectable credentials than the genre usually offers.

Elizondo writes, predictably, about the 2004 encounter between USS Nimitz pilots and a wingless white object with no visible control surfaces, the one Commander David Fravor nicknamed the "Tic Tac"; he writes about the later incidents involving the USS Theodore Roosevelt off the Virginia coast; he writes about pilots who, he claims, suffered unexplained physical effects after close encounters with these objects — a claim of considerable gravity, presented with the same narrative confidence he reserves for his recruitment stories at Fort Huachuca, with little apparent concern for the very different evidentiary weight the two kinds of episodes carry. This is a book that asks to be read, in one paragraph, as a matter of national security, and in the next, as a revelation about humanity's place in the cosmos, and these two ambitions do not reinforce each other so much as compete for the same narrative space — a competition the second almost always wins.

In fairness, Elizondo is not the most extravagant memoirist produced by the current cycle of public interest in UAP; he shows, compared with other figures in the genre, a certain verbal caution, generally avoiding the word "extraterrestrial" in favor of formulations like "non-human intelligence," a distinction the book presents as scientific rigor and a more suspicious reader might read as legal cover. Nor is he without narrative skill: the chapters on his training as a counterintelligence agent, though tangential to the book's central subject, are written with the craft of someone who has learned to structure a report to hold a busy superior's attention. The problem is not a lack of talent. The problem is that this talent is repeatedly put in service of erasing the distance between what the author knows, what the author believes, and what the author can prove — and the book depends on the reader not noticing the difference.

There is, moreover, a telling pattern in how "Imminent" treats its critics. Skeptics appear almost invariably as armchair analysts, cut off from classified material, unable to grasp the complexity of what Elizondo witnessed firsthand. It is a defense that works on the emotional register — who would not rather trust the eyewitness than the remote analyst? — but it dodges the simpler question: if classified evidence so clearly supports the book's claims, why has the partial declassification of recent years, including reports the Office of the Director of National Intelligence itself has delivered to Congress, failed to settle the matter? The book never turns that question on itself, and it is precisely the question it needs most.

It is worth placing "Imminent" within the genre that produced it, since the book does not arrive on empty ground. Leslie Kean published "UFOs: Generals, Pilots, and Government Officials Go on the Record" in 2010, a work of journalism that gathered testimony from military officers and officials using a methodology closer to reporting than to personal memoir, and one that remains, for many serious readers of the subject, the sturdier point of entry. Measured against that standard, "Imminent" is a step backward: it replaces the work of cross-verification with the authority of the first person, and asks the reader to accept that substitution as an improvement rather than a loss. It does not help that Elizondo's own account of the institutional lineage — the distinction between AAWSAP, the Bigelow-funded, congressionally financed program that ran from 2008 to 2010, and AATIP, the smaller, unfunded initiative that followed, which Elizondo himself describes as a kind of side project its members tended to "when their day jobs allowed" — took years to become public, and that the book itself only half clarifies it, leaving the reader to suspect that much of the public mythology around "the Pentagon's secret UFO program" rests on a conflation Elizondo has found it convenient not to correct too emphatically.

It is also worth noting the tone the author has chosen, because tone here functions as argumentative strategy as much as stylistic choice. Elizondo avoids histrionics; he does not write like an ecstatic convert but like a weary professional who would rather not have to tell any of this. That restraint operates as a form of indirect evidence — if the author does not need to exaggerate, perhaps the facts speak for themselves — and it is, at the same time, one of the oldest and most effective devices of persuasive nonfiction. Truman Capote used it to convince readers that "In Cold Blood" was pure reportage rather than novelized reconstruction; Elizondo uses it so that we forget how much of what he presents as verified fact is, at best, secondhand interpretation of data that no one outside a small circle can examine. The calm of the prose proves nothing, but it functions as if it did, and that may be the book's most genuine skill.

"Imminent" arrives at a moment when the UAP subject has stopped being marginal: congressional committees have held public hearings, whistleblowers with far less contested credentials than Elizondo's have testified under oath, and a growing share of the American public — a majority, according to recent polling — believes the government knows more than it says. In that context, a book like this one serves an almost inevitable function: not so much resolving the mystery as capitalizing on it. Elizondo writes as someone who already knows his testimony will be received with the reverence granted to the initiated, and that foreknowledge contaminates the honesty of the account; what is missing is discomfort, genuine doubt, any acknowledgment that the author's own institutional standing remains disputed by the very sources he invokes to vouch for his experience.

Closing the book leaves a feeling that is more melancholy than indignant — the sense of an opportunity missed. There is a better book buried inside "Imminent": one about how large bureaucracies manage uncertainty, about the temptation to mistake secrecy for authority, about what happens to an official who decides his personal conviction outweighs the institutional consensus that once employed him. That book would have been uncomfortable, nuanced, and, I suspect, far more persuasive than the one Elizondo ultimately wrote. Instead, "Imminent" chooses the promise of revelation over the discipline of evidence, and ends up resembling less an intelligence briefing than an act of faith drafted in the vocabulary of national security.

Rethinking Lunar Transportation Architecture

Rethinking Lunar Transportation Architecture:

A Critical Assessment of Orbital Refueling Dependencies and a Modular
Alternative

 

ABSTRACT

Current proposals for crewed lunar missions center on SpaceX's Starship vehicle, which requires between 8 and 16 orbital refueling operations per lunar sortie. While the architecture is technically feasible and economically motivated by vehicle reusability, it concentrates risk in an extended propellant-transfer chain whose cumulative failure probability is non-trivial. This paper critically evaluates the refueling-dependent architecture, identifies structural weaknesses frequently omitted in advocacy literature, and proposes a modular three-element alternative consisting of a reusable Earth-to-orbit transport, a permanent cislunar tug, and a dedicated lunar lander. The proposed architecture reduces compounded mission risk, improves long-term scalability, and is compatible with future in-situ resource utilization (ISRU). However, it entails significantly higher development costs and introduces its own set of operational complexities that must be acknowledged honestly. A phased roadmap for transition is presented, along with a comparative risk matrix. The authors conclude that neither architecture is categorically superior; the optimal path depends on mission cadence, budget horizon, and ISRU maturity.

Keywords: lunar architecture, orbital refueling, Starship, cislunar transport, ISRU, modular spacecraft design, mission reliability

 

1. Introduction

The return of humans to the lunar surface represents one of the most consequential engineering challenges of the twenty-first century. Unlike the Apollo program, which was sustained by extraordinary and politically contingent public expenditure, contemporary lunar ambitions must be commercially viable, operationally sustainable, and extensible toward deeper-space objectives.

SpaceX's Starship system has emerged as the centerpiece of NASA's Artemis crewed lunar landing strategy. As the designated Human Landing System (HLS), a modified Starship variant is expected to carry astronauts from lunar orbit to the surface and return them to the Gateway or directly to a crewed Orion capsule. To accomplish this, the vehicle must carry sufficient propellant for trans-lunar injection, orbital insertion, landing, ascent, and rendezvous — an energetically demanding sequence that, in a single vehicle, requires a propellant mass far exceeding what a single Starship can carry to orbit.

The proposed solution is orbital propellant transfer: multiple tanker Starships launch to low Earth orbit (LEO), transfer liquid methane (LCH₄) and liquid oxygen (LOX) to a depot or directly to the lunar Starship, and only then does the lunar vehicle depart. SpaceX estimates that between 8 and 16 tanker flights may be required per lunar mission, depending on propellant transfer efficiency and storage losses.

This paper does not dispute the technical feasibility of orbital refueling. It disputes the widespread failure to account for the compounded risk of requiring that many sequential operations to succeed before a single crewed mission can begin. Drawing on principles of systems-reliability engineering, mission architecture analysis, and historical spaceflight data, this paper presents a critical assessment of the current architecture and proposes a modular alternative that distributes risk more favorably across mission phases.

 

2. Critical Assessment of the Orbital Refueling Architecture

2.1 The Compounded Failure Probability Problem

In reliability engineering, systems in series — where all components must function for the system to succeed — have a total reliability equal to the product of each component's individual reliability. If each orbital refueling operation has a success probability of 0.99 (99%), ten such operations yield a cumulative success probability of 0.99¹⁰ ≈ 0.904. Sixteen operations yield 0.99¹⁶ ≈ 0.851. These figures are conservative; early-stage propellant transfer in cryogenic, microgravity conditions has not yet been demonstrated at the required scale.

SpaceX's demonstrated Starship reliability as of mid-2025 remains in early validation phases. Extrapolating that performance to 8–16 sequential tanker missions — each requiring precise rendezvous, docking, and cryogenic transfer — before a single crewed departure represents a significant and underacknowledged risk concentration.

2.2 Cryogenic Propellant Boil-Off and Storage Duration

Liquid oxygen and liquid methane are cryogenic propellants requiring storage at approximately -183°C and -161°C respectively. In the thermal environment of low Earth orbit, passive storage is insufficient; active cooling or zero-boil-off (ZBO) systems are required. NASA's technology development roadmaps acknowledge that ZBO in orbit remains an open engineering challenge at the scale required for Starship lunar missions.

If tanker missions are spread over days or weeks — as logistics realities suggest — propellant boil-off becomes a significant mass and cost penalty. The original article under review does not acknowledge this challenge. Neither does it engage with the engineering literature on long-duration cryogenic storage, which remains an active research problem.

2.3 The Displaced Complexity Fallacy

The article under review correctly identifies that the proposed three-vehicle modular architecture reduces orbital refueling requirements. However, it fails to acknowledge that the cislunar tug itself requires propellant, and that propellant for the tug must either come from Earth — reintroducing a supply chain problem — or from lunar ISRU, which is not yet operational. This is a case of displaced complexity: the problem is not eliminated; it is moved to a different subsystem and deferred to a later phase.

Furthermore, the proposed architecture requires three rendezvous events per mission: crew to tug in LEO, crew to lander in lunar orbit, and crew back to tug for return. The original architecture also requires rendezvous events. A rigorous comparison must account for the total rendezvous count and associated risk across both architectures, not merely the refueling events in isolation.

2.4 The Long-Duration Space Asset Problem

A cislunar tug permanently stationed in space introduces a category of engineering challenge absent from the original critique: long-duration exposure to the space environment. Galactic cosmic radiation, solar energetic particles, micrometeoroids, and the thermal cycling of deep space progressively degrade structural materials, electronics, and propellant management systems. The International Space Station, by comparison, requires extensive and continuous maintenance by resident crews.

A permanently stationed tug that is never returned to Earth for servicing must either be designed for autonomous maintenance — an undemonstrated capability at this scale — or be serviced in situ, which requires crewed maintenance missions and a robust spare parts supply chain. Neither the original article nor most modular architecture proposals address this in detail.

2.5 Development Cost and Industrial Base

SpaceX's vertical integration strategy — using one vehicle platform (Starship) for Earth launch, tanker, and lunar descent — dramatically compresses the development cost curve. The learning-by-flying approach has already demonstrated rapid iteration on Starship's design. Developing three separate vehicle classes, each requiring independent testing, certification, qualification, and operational support infrastructure, is substantially more expensive in the near term.

For a commercially driven program, the near-term cost structure matters enormously. A modular architecture that is theoretically superior in steady-state operations but costs 3–5× more to develop may not survive the political and financial constraints of the 2020s and 2030s space budget environment.

 

3. A Realistic Modular Architecture Proposal

3.1 Foundational Principles

A credible modular lunar architecture must satisfy four criteria simultaneously: (1) demonstrably lower compounded mission risk than the baseline architecture; (2) a realistic development and deployment timeline compatible with existing budgets and industrial capacity; (3) compatibility with near-term ISRU absence; and (4) a clear evolutionary path toward long-term sustainability. Proposals that satisfy only future-state criteria while ignoring present-state constraints are analytically incomplete.

3.2 Element 1 — Reusable Earth-to-Orbit Transport

The first element is a reusable launch vehicle optimized for Earth ascent and descent only. Starship itself, in its standard non-lunar configuration, is a strong candidate. Its role is restricted to crew and cargo delivery to a designated orbital node — Gateway, a dedicated propellant depot, or a high-altitude parking orbit. It does not transit to the Moon, does not serve as a propellant tanker, and does not interface with lunar operations.

By restricting its role, the vehicle can be iteratively improved against a single performance metric: low-cost, high-cadence access to orbit. This aligns with SpaceX's existing development trajectory and requires no new development program.

3.3 Element 2 — Permanent Cislunar Transfer Vehicle

The second element is a transfer vehicle permanently stationed in cislunar space, operating between LEO or the Gateway node and a low lunar orbit (LLO) staging point. Because it never enters Earth's atmosphere, it requires no thermal protection system, no aerodynamic control surfaces, and no atmospheric entry-rated structures. Its mass budget is therefore dedicated entirely to propulsion, crew habitation for transit, and rendezvous/docking systems.

Critically, this vehicle's propellant supply chain must be specified honestly. In the near term (Phase 1–2), it is supplied by tanker launches from Earth — but to cislunar orbit, not to LEO. The advantage is that fewer tanker missions are needed per crew transit, because the tug's mass and propellant budget are significantly lower than a full lunar Starship's. In the long term (Phase 3+), it is refueled by lunar ISRU-derived propellant.

The tug architecture aligns with elements already under development: NASA's Gateway Power and Propulsion Element uses solar electric propulsion for cislunar orbital maintenance. A chemical-propulsion crew transfer variant would complement, not duplicate, this capability.

3.4 Element 3 — Dedicated Lunar Lander

The third element is a lander optimized exclusively for the lunar environment: vacuum descent, surface operations, and ascent to LLO. It carries no Earth-return propellant, no reentry systems, and no atmospheric interface hardware. Its design envelope is narrower than a multi-environment vehicle, which permits significant mass reduction.

The lander may be reusable within the lunar system (descending and ascending multiple times before requiring maintenance) or may be partially expendable at the ascent stage level — following the Apollo Lunar Module model. The choice depends on mission cadence. For fewer than four missions per year, a partially expendable lander is likely more cost-effective than developing and maintaining a fully reusable one.

3.5 ISRU: Necessary Condition, Not Near-Term Reality

In-situ resource utilization — the extraction of water ice from permanently shadowed regions (PSRs) at the lunar poles, its electrolysis into liquid hydrogen and oxygen, and its liquefaction for propellant use — is the enabling technology for long-term modular architecture sustainability. It is not, however, an enabling technology for near-term missions.

NASA's MOXIE experiment on the Perseverance rover demonstrated oxygen production from Martian atmospheric CO₂ at a rate of approximately 6 grams per hour. Scaling this to the tonnes of propellant required for a cislunar tug represents an engineering challenge many orders of magnitude larger. Current estimates suggest that a pilot-scale ISRU plant capable of producing commercially meaningful propellant quantities on the Moon is a 2038–2045 prospect under optimistic assumptions.

Any modular architecture proposal that requires ISRU to function must therefore also specify a credible bridging strategy for the pre-ISRU era. This paper proposes that Earth-supplied propellant delivered to a cislunar depot — not to LEO — serves this bridging function.

 

4. Comparative Risk and Performance Analysis

Table 1 presents a comparative matrix of the Starship-centric architecture and the proposed modular alternative across eight mission-critical criteria. Assessments are qualitative but grounded in the engineering literature cited throughout this paper.

 

 

 

 

 

 

 

 

The matrix reveals that neither architecture is universally superior. The Starship approach offers decisive advantages in near-term development cost and ISRU-independence. The modular approach offers meaningful advantages in compounded mission risk, long-term scalability, and infrastructure growth. The optimal architecture is therefore contingent on mission objectives, budget horizon, and the rate of ISRU technology maturation.

 

5. Phased Implementation Roadmap

A realistic transition from the current architecture toward a mature modular system requires a phased approach that does not abandon near-term operational capability in pursuit of long-term optimality. Table 2 presents a four-phase roadmap with associated milestones and critical risks.

 

Phase 1 is explicitly designed to generate operational experience with the Starship architecture rather than replace it. Data from early Artemis missions will provide ground truth on orbital refueling performance, propellant boil-off rates, and cislunar rendezvous reliability — all of which are currently estimated parameters. Architectural decisions for Phase 2 and beyond should be informed by this empirical data rather than by pre-mission assumptions.

 

6. Discussion

The broader literature on lunar architecture consistently underestimates the degree to which architecture choices interact with programmatic, financial, and geopolitical constraints. The modular architecture proposed in this paper is technically more robust than a single-vehicle approach in the steady-state, high-cadence operational regime. It is not more robust in the near-term development and initial operations regime.

Two observations deserve particular emphasis. First, the Gateway lunar station — already under development by NASA and its international partners — represents a nascent cislunar node that is architecturally compatible with the tug concept proposed here. Rather than treating Gateway as an unnecessary complexity, as some commercial advocates do, a modular architecture framework recontextualizes it as a foundational infrastructure element. This has implications for international partnership structures and cost-sharing arrangements.

Second, the geopolitical dimension of ISRU access is entirely absent from most technical architecture discussions, including the article under review. The water ice deposits in PSRs near the lunar south pole are not uniformly distributed, and access to the highest-concentration sites will be contested. The Artemis Accords establish a framework for peaceful ISRU activities but do not resolve priority disputes. An architecture that depends on ISRU for its long-term propellant supply is also an architecture that depends on sustained, uncontested access to specific lunar geographic locations — a geopolitical assumption that merits explicit treatment.

 

7. Conclusions

This paper has critically examined the SpaceX Starship orbital refueling architecture for crewed lunar missions and proposed a modular three-element alternative. The key conclusions are as follows:

(1) The orbital refueling architecture concentrates compounded mission risk in a propellant-transfer chain that has not yet been demonstrated at operational scale. Cumulative failure probability across 8–16 tanker missions is non-trivial and is systematically underreported in advocacy literature.

(2) The modular alternative — comprising a reusable Earth-to-orbit transport, a permanent cislunar transfer vehicle, and a dedicated lunar lander — distributes risk more favorably across mission phases and offers superior long-term scalability. However, it entails substantially higher development costs and introduces its own operational complexities, most notably the long-duration maintenance of assets permanently stationed in space.

(3) In-situ resource utilization is a necessary condition for the long-term viability of any lunar architecture but is not a near-term engineering reality. Proposals that present ISRU as a near-term solution are analytically premature. A credible modular architecture must specify a viable bridging propellant strategy for the pre-ISRU era.

(4) Neither architecture is categorically superior. The Starship-centric approach is appropriate for the near term, provided that operational data from early missions is used to rigorously evaluate the refueling chain's actual performance. The modular approach becomes progressively more attractive as mission cadence increases and ISRU technology matures.

(5) The geopolitical dimension of ISRU access to lunar polar resources is a first-order strategic variable that technical architecture analyses routinely omit. It must be incorporated into any serious long-term lunar transportation planning.

 

References

1 Colaprete et al. (2010) — LCROSS water detection, Science

2 Drake, B.G. (2009) — Mars DRA 5.0, NASA SP-2009-566 

3 Hoffman, Hecht et al. (2022) — MOXIE, Science Advances, DOI real 

4 Metzger et al. (2013) — Space Bootstrapping, J. Aerospace Eng. 

5 NASA (2022) — National Cislunar Strategy, White House

6 NASA (2022) — NRHO Artemis Architecture, Official White Paper

7 Sanders & Larson (2013) — ISRU integration, Advances in Space Research