# Learning Curves for Onboard Autonomy: Does Each Successive Autonomous-Operations Flight Demonstration Lower the Cost-to-Field of the Next?

**Candidate:** JPL_AUTONOMY_EDL_01
**Program:** COLLEGIUM 1st Battalion
**North Star / JPL category:** Autonomous Systems and Robotics
**Method:** Quantitative (Wright/Henderson log-log experience-curve regression with fixed effects)
**Date:** 2026-06-15

---

## Abstract

The cost of qualifying onboard spacecraft autonomy for flight is widely assumed to fall as the technology matures, but this assumption has not been tested as a measurable experience-curve relationship. This dissertation specifies and operationalizes a test of whether the recurring engineering cost to qualify a class of onboard autonomy capability declines as a power-law function of cumulative flight-demonstrated heritage. The unit of analysis is the capability-class development episode: a mission or technology demonstration that fields a defined autonomy capability such as onboard planning, autonomous science target selection, autonomous navigation, or autonomous entry-descent-and-landing hazard handling. The contribution is a single falsifiable claim. The null hypothesis (H0) is that per-episode autonomy development cost is flat with respect to the cumulative count of prior flight demonstrations within a capability class. The alternative hypothesis (H1) is that per-episode cost declines along a log-log experience curve with a statistically detectable negative slope after controlling for capability class and decade. The estimator is an ordinary-least-squares regression of the logarithm of normalized development cost on the logarithm of cumulative flight-demonstrated heritage, with capability-class and decade fixed effects, fit to a constructed panel assembled from NASA TechPort project records and technology-readiness-level histories, NTRS autonomy demonstration reports for Deep Space 1 Remote Agent, the EO-1 Autonomous Sciencecraft Experiment, AEGIS, and Ingenuity, and NASA Instrument Cost Model-class parametric development-cost estimates. The work is presented at the design stage. The data construction, model specification, identification strategy, and threats to validity are fully developed, and an illustrative, explicitly labeled non-empirical specification of expected results is provided. No fitted coefficients from the full dataset are reported, because the panel assembly and cost normalization are not yet complete. The intended outcome is a defensible measurement of the autonomy experience-curve slope, or a credible failure to reject the flat-cost null, either of which informs NASA and JPL technology-investment sequencing.

---

## 1. Introduction and Contribution

### 1.1 The problem

NASA and the Jet Propulsion Laboratory invest in onboard autonomy because autonomy lifts the science and operational return of missions that operate under long communication delays and constrained ground contact. The Deep Space 1 Remote Agent Experiment in 1999 was the first time an artificial-intelligence agent controlled a NASA spacecraft (NTRS 20210003369). The EO-1 Autonomous Sciencecraft Experiment moved onboard event detection and replanning into routine Earth-observation operations after 2003 (Chien et al. 2005a, doi:10.2514/1.12923). AEGIS placed autonomous science target selection first on the Opportunity rover and then on the Curiosity ChemCam instrument (Estlin et al. 2012, doi:10.1145/2168752.2168764; Francis et al. 2017, doi:10.1126/scirobotics.aan4582). The Mars 2020 mission flew autonomous navigation at scale on Perseverance and demonstrated powered autonomous flight with the Ingenuity helicopter (Verma et al. 2023, doi:10.1126/scirobotics.adi3099; Balaram et al. 2021, doi:10.1007/s11214-021-00815-w).

Each of these episodes is treated inside NASA as a heritage asset. The working assumption is that once a capability has flown, the next mission that needs that capability can field it more cheaply, because the design patterns, verification approaches, and flight-software components are reusable. This assumption drives portfolio sequencing decisions and technology-readiness-level accounting. Yet the assumption is rarely stated as a measurable quantity. The question of how much cheaper the next demonstration becomes, and whether the cost decline follows a regular curve, has not been answered with a fitted model on a constructed dataset.

### 1.2 The gap in the literature

Two literatures are relevant and have not been joined. The first is the experience-curve literature in the economics of technology. Since the airframe cost-quantity studies of the 1930s and 1950s, unit production cost has been observed to fall as a power-law function of cumulative output across many manufactured goods (Thompson 2012, doi:10.1257/jep.26.3.203). Modern treatments formalize the experience curve as a statistical forecasting object with quantifiable error (Nagy et al. 2013, doi:10.1371/journal.pone.0052669; Farmer and Lafond 2016, doi:10.1016/j.respol.2015.11.001; Lafond et al. 2018, doi:10.1016/j.techfore.2017.11.001). This literature is almost entirely about manufactured hardware unit cost, not about the non-recurring engineering cost of qualifying a software-intensive capability for flight.

The second literature is the spacecraft-autonomy and technology-maturation literature. It documents individual autonomy demonstrations in detail and reviews the state of space autonomy (Gao and Chien 2021, doi:10.1007/s43154-021-00057-2), and it analyzes the technology-readiness-level construct used to track maturation (Mankins 2009, doi:10.1016/j.actaastro.2009.03.058; Olechowski et al. 2020, doi:10.1002/sys.21533). This literature describes heritage qualitatively but does not estimate whether heritage lowers cost along a measurable curve.

The gap is the absence of an experience-curve treatment of onboard autonomy qualification cost. No published study fits a Wright-type log-log model of capability-class autonomy development cost on cumulative flight-demonstrated heritage.

### 1.3 The single falsifiable contribution

This dissertation tests one claim.

- **H0 (null):** Per-episode autonomy development cost is flat with respect to cumulative prior flight demonstrations within a capability class. The experience-curve slope coefficient is statistically indistinguishable from zero.
- **H1 (alternative):** Per-episode autonomy development cost declines along a log-log experience curve as cumulative flight-demonstrated heritage accumulates. The slope coefficient is negative and statistically significant after controlling for capability class and decade.

The contribution is the measurement itself: a fitted slope, a confidence interval, and an explicit accept-or-reject decision on H0. The contribution is falsifiable because a fitted slope that is zero or positive, or that is not statistically distinguishable from zero, rejects H1 and supports the flat-cost null.

### 1.4 Why it matters for NASA and JPL

If autonomy qualification cost falls along a measurable curve, the slope is a planning parameter. It tells a program office how much heritage investment is required before a target capability becomes affordable for a cost-capped mission class, and it supports sequencing decisions in the Autonomous Systems and Robotics portfolio. A concrete use is the build-or-wait decision: when a future mission needs a capability that is currently too expensive to qualify within its cost cap, a known experience-curve slope tells the portfolio whether an intermediate, lower-stakes demonstration would buy down enough cost to bring the target capability within reach, and roughly how many such demonstrations would be required. Without a measured slope, that decision is made by assertion. If the cost is flat, the heritage-reuse argument that justifies much autonomy investment is weaker than assumed, and the portfolio rationale must rest on capability value rather than on declining cost. Either outcome is decision-relevant. The measurement also gives a quantitative complement to the technology-readiness-level scale, which tracks maturity but not the cost of advancing it, and which is known to be ordinal and non-monetary (Olechowski et al. 2020, doi:10.1002/sys.21533). A cost-anchored learning rate would let a program reason about the dollars, not only the readiness level, attached to advancing a capability across the demonstration sequence.

---

## 2. Background and Literature

### 2.1 The experience curve as a measurement object

The experience curve, in its Wright form, states that the cost per unit falls by a constant percentage each time cumulative output doubles. Written in logarithms, the relationship is linear: the logarithm of unit cost is a linear function of the logarithm of cumulative output, and the slope of that line encodes the learning rate. The airframe studies established the empirical regularity for aircraft production, and Thompson (2012, doi:10.1257/jep.26.3.203) reviews the long history of the cost-quantity relationship and the evidence that organizational learning-by-doing, not only static scale economies, drives it. The modern econometric literature treats the experience curve as a forecasting model and asks how reliably it predicts. Nagy et al. (2013, doi:10.1371/journal.pone.0052669) compare functional forms across many technologies and find the Wright power law performs well. Farmer and Lafond (2016, doi:10.1016/j.respol.2015.11.001) quantify the predictive accuracy of experience curves and characterize their error distribution, and Lafond et al. (2018, doi:10.1016/j.techfore.2017.11.001) extend this to distributional forecasts. The Indian cost-quantity evidence (Bhattacharya et al. 2017, doi:10.1016/j.respol.2017.09.005) documents that learning rates are heterogeneous across settings, which motivates the use of fixed effects when pooling heterogeneous capability classes.

The lesson this dissertation takes from the experience-curve literature is methodological. The log-log specification is the standard and best-validated functional form, the learning rate is the object of interest, and the central empirical risk is that observed cost declines reflect confounders such as scale, time trends, or input-price changes rather than cumulative experience.

A further methodological point from this literature is the distinction between the Wright formulation, in which cost is a function of cumulative output, and the Moore formulation, in which cost is a function of calendar time. Nagy et al. (2013, doi:10.1371/journal.pone.0052669) show that where output grows exponentially the two are observationally similar, which is why a credible experience-curve study must include a time control to separate genuine learning from a coincident time trend. This study adopts that discipline directly through decade fixed effects, so that the estimated slope is identified from the part of cumulative heritage that does not move mechanically with calendar time. The literature also warns that the dependent variable must be a true unit cost or a true per-episode cost rather than a total that grows with scale, which is why the present design normalizes development cost by capability scope before taking logarithms. Finally, the experience-curve tradition treats the learning rate as a transferable parameter only within a coherent technology family, which is the formal justification for organizing the autonomy panel into capability classes and estimating a within-class slope rather than a single pooled slope across heterogeneous capabilities.

### 2.2 The anchor frameworks in plain language

This work applies two methodological lenses from the Hall of Shoulders.

**Joel Mokyr: useful knowledge and the conditions for cumulative cost decline.** Mokyr distinguishes propositional knowledge, which is knowledge of why something works, from prescriptive knowledge, which is knowledge of how to do something. His central argument is that a technique becomes cheap to reproduce and extend only when it rests on a wide base of propositional understanding; techniques discovered by trial without underlying theory tend to stagnate rather than improve (Mokyr, *The Gifts of Athena*, 2002). Applied here, Mokyr predicts that the autonomy capability classes most likely to show a steep cost-decline curve are those whose underlying propositional base, for example the formal theory of automated planning, estimation, and verification, is mature enough that each demonstration codifies reusable, self-correcting knowledge rather than a bespoke artifact. Mokyr also stresses that progress is reversible and that access cost to knowledge governs diffusion. This predicts that the experience-curve slope will be steeper where the heritage knowledge is openly codified and reusable, for example through the Core Flight System and shared autonomy frameworks, and flatter where each project re-implements from scratch. Mokyr's lens therefore supplies a testable moderator: codification and reuse infrastructure should be associated with steeper cost decline.

**W. Brian Arthur: increasing returns and learning effects.** Arthur's theory of increasing returns identifies learning effects as one of the four mechanisms that make a technology more attractive the more it is adopted (Arthur 1989, doi:10.2307/2234208; Arthur 1994, doi:10.3998/mpub.10029). In Arthur's framework, learning effects are precisely the mechanism that an experience curve measures: each use of a technique lowers the cost of the next use, generating positive feedback. Arthur also warns that increasing-returns systems are path-dependent and non-ergodic, so early choices lock in and the realized cost path reflects historical sequence rather than only technical fundamentals (Arthur 2021, doi:10.1038/s42254-020-00273-3). Applied here, Arthur predicts that a measurable negative experience-curve slope is the expected signature if learning effects dominate, but he also cautions that the realized slope is contingent on which capability classes received early investment. The launch-cadence literature gives a concrete space example of his learning-effect mechanism, where each flight lowers the cost of the next (CSET 2024, doi:10.51593/20240017; IAC 2024, doi:10.52202/078367-0036). Arthur's lens therefore supports H1 as the theoretically expected direction while flagging path dependence as a threat to clean identification.

### 2.3 The autonomy demonstration record

The autonomy demonstrations that form the empirical spine of this study are documented in the primary literature, and their chronology is the source of the cumulative-heritage variable. The Remote Agent Experiment integrated onboard planning, execution, and fault recovery and was validated and verified as a flight experiment in 1999 (NTRS 20210003369; NTRS 20210001679). It is the origin observation for the onboard planning and scheduling capability class and for the autonomous fault detection and recovery class. The lessons-learned record from Deep Space 1 documents the organizational cost of infusing autonomy into a flight project and reports that the impact of inserting system-level autonomy into a flight project was a major surprise to the project (NTRS 20210003565). That record is direct primary evidence on the non-recurring engineering and organizational cost of a first-of-kind autonomy capability, which is the quantity the dependent variable is intended to capture.

The EO-1 Autonomous Sciencecraft Experiment moved onboard detection and replanning into operational Earth observation after 2003 and improved the quality and timeliness of returned science data (Chien et al. 2005a, doi:10.2514/1.12923). It was used operationally to detect and monitor active volcanism, floods, and cryospheric change, with onboard detection triggering autonomous retasking (Chien et al. 2005b, doi:10.1016/j.rse.2005.08.007). EO-1 is the transition point at which onboard autonomy moved from a bounded experiment to a routine operational element, and it is the second observation in the onboard planning and scheduling class.

AEGIS established autonomous science target selection first on the Opportunity rover in 2010 and then on the Curiosity ChemCam instrument in 2016, reducing the operations latency imposed by ground-in-the-loop targeting of narrow-field instruments (Estlin et al. 2012, doi:10.1145/2168752.2168764; Francis et al. 2017, doi:10.1126/scirobotics.aan4582). The AEGIS sequence is the cleanest within-class heritage pair in the record: the same capability was reused on a second platform, and the second deployment is documented as building on the first. It is the central illustrative case for the heritage-lowers-cost mechanism the study tests.

Perseverance flew autonomous navigation across most of its traverse, with the AutoNav system evaluating the large majority of the distance driven in early operations, and Ingenuity demonstrated autonomous powered flight as a technology demonstration in 2021 (Verma et al. 2023, doi:10.1126/scirobotics.adi3099; Balaram et al. 2021, doi:10.1007/s11214-021-00815-w). These are the most recent and most capable observations in the autonomous navigation class. The space-autonomy review by Gao and Chien (2021, doi:10.1007/s43154-021-00057-2) situates these demonstrations within a longer arc and states the recurring claim, that flown heritage lowers the barrier to the next demonstration, that this study converts into a testable hypothesis. The Core Flight System community work documents the reusable flight-software substrate that the heritage-reuse argument depends on and that Mokyr's framework would identify as the codified knowledge base whose presence should steepen the cost-decline curve (cFS community, NTRS, 2016).

### 2.4 Cost measurement for space development

The development cost of space systems is estimated parametrically. Single-variable parametric cost models for space telescopes (Stahl et al. 2010, doi:10.1117/1.3456582) exemplify the NASA Instrument Cost Model family of estimators that this study uses to normalize and cross-check development costs. The technology-readiness-level construct is the standard maturity tracker (Mankins 2009, doi:10.1016/j.actaastro.2009.03.058), and its known shortcomings, including its ordinal and non-cost character, are documented (Olechowski et al. 2020, doi:10.1002/sys.21533). These shortcomings are why this study treats the technology-readiness-level history as a maturation covariate, not as a cost proxy.

---

## 3. Data

### 3.1 Named datasets and sources

The study assembles one panel from four named sources.

1. **NASA TechPort project records and technology-readiness-level histories.** TechPort is NASA's system of record for technology projects. It provides project descriptions, start and end dates, organizational ownership, technology-readiness-level entry and exit estimates, and taxonomy classification. Access path: the TechPort public application programming interface and bulk data export at the TechPort public portal. TechPort supplies the project inventory, the capability-class taxonomy mapping, the maturation covariate, and project timing.

2. **NTRS autonomy demonstration reports.** The NASA Technical Reports Server provides the primary engineering and lessons-learned documentation for the named demonstrations: Remote Agent on Deep Space 1 (NTRS 20210003369; 20210001679; 20210003565), the EO-1 Autonomous Sciencecraft Experiment, AEGIS, and Ingenuity. Access path: the NTRS citations search application programming interface. NTRS supplies the heritage chronology, the capability definitions, and qualitative scope information used to size each episode.

3. **NICM-class development-cost estimates.** NASA Instrument Cost Model-class parametric estimates and the published parametric cost-model literature (Stahl et al. 2010, doi:10.1117/1.3456582) supply normalized development-cost estimates and the cost-normalization basis. Where project-specific costs are not public, NICM-class parametric estimates provide a consistent imputation and a cross-check on reported figures.

4. **Published mission and autonomy literature.** The peer-reviewed record listed in Section 2 supplies capability scope, demonstration dates, and the heritage links between episodes.

### 3.2 Unit of analysis

The unit of analysis is the capability-class development episode. An episode is a single mission or technology demonstration that fields a defined onboard autonomy capability. A capability class is a coherent functional category of onboard autonomy. The initial capability classes are: onboard planning and scheduling, autonomous science target selection, autonomous surface or in-flight navigation, autonomous fault detection and recovery, and autonomous entry-descent-and-landing hazard handling. Each episode is assigned to exactly one primary capability class. Episodes that field more than one capability are split or assigned to their dominant capability with a flag.

### 3.3 Variable construction

- **Dependent variable, normalized development cost.** The recurring engineering cost to qualify the capability for flight, expressed in constant-year dollars, normalized by capability scope using a NICM-class basis. Normalization removes the part of cost driven by instrument or platform scale so that the residual reflects the autonomy qualification effort. The variable enters the model in natural logarithm.

- **Independent variable, cumulative flight-demonstrated heritage.** For each episode, the count of prior flight demonstrations within the same capability class that completed before the episode's development start date. This is the experience stock. It enters the model in natural logarithm, with a standard treatment of the first observation in each class, for which the cumulative count is set to one before taking the logarithm so the first demonstration is the curve origin.

- **Capability-class fixed effects.** Indicator variables for each capability class, absorbing time-invariant differences in baseline cost and intrinsic difficulty across classes.

- **Decade fixed effects.** Indicator variables for the decade of development start, absorbing economy-wide and agency-wide time trends in cost, tooling, and computing that affect all classes.

- **Maturation covariate.** The technology-readiness-level at development start, from TechPort, included as a robustness covariate to separate the heritage effect from the simpler effect of starting at a higher maturity.

The construction of the dependent variable is the most delicate step and deserves a fuller statement. Public development-cost figures for autonomy capabilities are reported inconsistently: some are full project costs that bundle the platform, instrument, and autonomy together; some are subsystem costs; and some are not disclosed. The normalization proceeds in three layers. The first layer extracts, where possible, the autonomy-specific non-recurring engineering portion from project documentation, separating it from platform and instrument cost. The second layer, where the autonomy portion is not separately reported, applies a NICM-class parametric estimate of the autonomy development effort given the documented scope, using the parametric cost-model approach exemplified by Stahl et al. (2010, doi:10.1117/1.3456582). The third layer expresses every figure in constant-year dollars using a standard deflator and records, for each observation, which layer produced the figure and an associated reliability flag. The reliability flag is then used in the weighted robustness specification so that observations resting on the weakest imputation receive less weight. This layered procedure does not eliminate measurement error in the dependent variable, but it makes the error explicit and auditable, which is the standard a defensible cost study must meet.

The cumulative-heritage variable also requires a stated convention for what counts as a completed prior demonstration. The convention adopted here is that a prior demonstration counts toward the heritage stock of a later episode only if it reached flight operation before the later episode's development start date, because heritage that arrives after development has begun cannot have lowered that development's cost. This forward-only rule prevents a mechanical correlation in which later, more expensive episodes are credited with heritage they could not have used. The rule is conservative: it tends to undercount heritage and therefore biases the estimated slope toward zero, making any rejection of the flat-cost null more credible.

### 3.4 Coverage

The intended coverage is autonomy flight demonstrations from the Deep Space 1 Remote Agent Experiment in 1999 through the Mars 2020 demonstrations and their immediate successors. This window spans the period in which onboard autonomy moved from single experiments to routine mission elements, which is the period over which an experience curve, if it exists, should be detectable.

### 3.5 Limitations

The data have four material limitations. First, the population of flight-demonstrated autonomy episodes is small, on the order of tens, not thousands, which limits statistical power and constrains the number of fixed effects that can be estimated. Second, development-cost figures are heterogeneous in definition and are sometimes not public, which forces NICM-class imputation and introduces measurement error in the dependent variable. Third, capability-class assignment requires judgment and can be contested at the boundaries. Fourth, heritage is not always within-agency or within-class; software components and design patterns cross capability classes, which the within-class heritage count does not capture and which biases against finding an effect. These limitations are addressed in the threats-to-validity analysis and bound the strength of any conclusion.

---

## 4. Research Design and Identification

### 4.1 Estimator

The estimator is ordinary least squares applied to a log-log experience-curve specification with two-way fixed effects. The estimating equation for episode i in capability class c and decade d is:

ln(Cost_icd) = alpha + beta * ln(CumHeritage_icd) + gamma_c + delta_d + epsilon_icd

where Cost_icd is normalized development cost, CumHeritage_icd is cumulative within-class flight-demonstrated heritage, gamma_c are capability-class fixed effects, delta_d are decade fixed effects, and epsilon_icd is the error term. The coefficient beta is the experience-curve slope and is the single parameter on which the hypotheses turn. A negative and statistically significant beta supports H1. A beta indistinguishable from zero supports H0. The implied learning rate is one minus two raised to the power beta, reported for interpretability.

The choice of ordinary least squares on the log-log form, rather than a non-linear or Bayesian estimator, is deliberate and follows the experience-curve literature. The log-log linearization is the form whose predictive performance has been validated across many technologies (Nagy et al. 2013; Farmer and Lafond 2016), it produces a single interpretable slope, and it keeps the degrees-of-freedom cost low, which matters acutely in a small panel. The fixed effects are entered as indicator variables rather than as a random-effects structure, because the capability classes and decades are the specific, non-sampled categories of interest rather than draws from a larger population, and because a fixed-effects estimator does not require the orthogonality assumption between the heritage variable and the class or decade effects that a random-effects estimator would impose. The trade-off is that fixed effects consume degrees of freedom; the analysis plan therefore caps the number of capability classes at the level the sample can support and reports the realized residual degrees of freedom alongside every coefficient.

### 4.2 Identification strategy

Identification of beta rests on within-capability-class, within-decade variation in cumulative heritage. The capability-class fixed effects remove the concern that some classes are simply cheaper than others for reasons unrelated to heritage. The decade fixed effects remove the concern that costs fell over time for reasons common to all classes, such as cheaper computing or improved software tooling. After these are absorbed, beta is identified from the comparison of episodes within the same class that faced different accumulated heritage at different points within the same decade. The maturation covariate provides a further check that beta is not merely capturing the tendency of later projects to start at higher technology-readiness levels.

### 4.3 Variables and specification choices

The baseline specification uses the within-class heritage count. Three robustness specifications are pre-registered in the analysis plan: a specification adding the technology-readiness-level maturation covariate; a specification that broadens heritage to include cross-class software-component reuse, which should attenuate any downward bias from shared components; and a specification that weights observations by the inverse of the imputation error to down-weight episodes with the least reliable cost figures.

### 4.4 Threats to validity

**Internal validity.** The primary internal threat is omitted-variable confounding between cumulative heritage and the secular decline in computing and software cost. Decade fixed effects mitigate but do not eliminate this; a within-decade computing trend could remain. A second internal threat is reverse pathways, where cheaper episodes are attempted only after heritage exists, which would make the heritage-cost association partly a selection artifact. The maturation covariate and the cross-class heritage robustness specification address this.

**External validity.** The findings, if any, generalize to NASA and JPL onboard autonomy for deep-space and planetary missions. They may not generalize to commercial autonomy, to terrestrial autonomy, or to mission classes outside the demonstration window. The fixed-effects design estimates an average within-class slope and does not claim a universal learning rate.

**Construct validity.** The dependent variable, autonomy qualification cost, is a construct that public records measure imperfectly, because cost-accounting boundaries differ across projects. The NICM-class normalization is itself a construct choice. The heritage count is a coarse measure of the true reusable-knowledge stock that Mokyr's framework identifies as the real driver. The cross-class robustness specification partly addresses the construct gap between the heritage count and reusable knowledge.

**Statistical-conclusion validity.** The small sample is the dominant statistical threat. With tens of observations and two sets of fixed effects, degrees of freedom are scarce and standard errors are wide. The analysis plan therefore pre-commits to reporting confidence intervals, to using small-sample-robust inference, and to treating a wide interval that contains zero as a failure to reject H0 rather than as evidence for H1. Multiple-specification testing is controlled by pre-registering the robustness specifications rather than searching over them.

---

## 5. Analysis Plan and Findings

### 5.1 Status

This section is a design-stage analysis plan. The full panel has not yet been assembled, the cost normalization is not yet complete, and no coefficients have been fitted to the full dataset. The numbers in Section 5.4 are illustrative and are labeled as such. They are not empirical results. They exist only to show what a result would look like and how it would be read against the hypotheses.

### 5.2 Estimation procedure

The estimation will proceed in five steps. First, build the episode inventory from TechPort and NTRS, assigning each episode to a capability class and recording its development-start date. Second, construct the cumulative within-class heritage count for each episode from the dated demonstration record, applying the forward-only counting rule. Third, assemble normalized development cost from public figures where available and NICM-class parametric estimates where not, recording the imputation layer and a reliability flag for each. Fourth, fit the two-way fixed-effects log-log regression and the three pre-registered robustness specifications. Fifth, report beta, its confidence interval, the implied learning rate, and the accept-or-reject decision on H0, together with diagnostic plots of log cost against log heritage by class.

Two pre-analysis checks are required before the slope is interpreted. The first is a check on the fixed-effects feasibility: if any capability class or any decade contains only a single observation, that cell contributes nothing to the within estimator and is reported as such, and the affected observations are flagged so the reader knows the slope is identified off a smaller effective sample than the raw count suggests. The second is an influence diagnostic: because the panel is small, a single high-leverage episode, for example an unusually expensive first-of-kind demonstration, could drive the slope. The plan therefore reports the slope with and without each potentially influential observation and treats a slope that depends on a single episode as not robust. These checks are stated in advance so that they constrain interpretation rather than being selected after seeing the result. The estimation code, the assembled panel, and the imputation log will be retained so the measurement is reproducible and auditable, consistent with the standard that a cost study must be checkable by a reviewer who did not build it.

### 5.3 Decision rule

H0 is rejected in favor of H1 if and only if the estimated beta is negative and its confidence interval excludes zero in the baseline specification and in at least two of the three robustness specifications. If beta is not distinguishable from zero, or is positive, H0 is not rejected and the flat-cost conclusion stands. This rule is fixed before estimation.

### 5.4 Illustrative, non-empirical expected results

The following figures are illustrative only and are not derived from a fitted model. They show the shape of a result that would reject H0. If onboard autonomy qualification cost followed an experience curve with a learning rate in the range commonly observed for software-intensive technologies, the slope beta would be a moderate negative number, the implied cost reduction per doubling of cumulative heritage would be a meaningful fraction, and the confidence interval would exclude zero. A plausible illustrative reading would be a negative slope with an implied per-doubling cost reduction somewhere in the low tens of percent, consistent in direction with Arthur's learning-effect mechanism and steeper for capability classes with mature, codified knowledge bases as Mokyr's framework predicts. Equally, the analysis may return a slope indistinguishable from zero, which would not reject H0 and would indicate that within-class heritage, as measured, does not lower qualification cost on its own. Both outcomes are reportable and both are informative. No claim is made here about which will occur.

### 5.5 What a falsifying result looks like

The contribution is falsified if, on the assembled panel, the estimated beta is zero or positive, or its confidence interval contains zero across specifications. In that case the heritage-lowers-cost assumption is not supported as an experience-curve relationship, and the study reports the flat-cost null as its finding.

---

## 6. Discussion

### 6.1 Implications

If H1 holds, the fitted slope is a usable planning parameter for the Autonomous Systems and Robotics portfolio. It quantifies how much heritage investment lowers the cost of the next demonstration and supports sequencing decisions, for example whether to invest in an intermediate demonstration to bring a target capability within a cost cap. If H0 holds, the portfolio rationale shifts from cost reduction to capability value, and the agency should not rely on heritage-driven cost decline to justify autonomy investment. The result also gives a cost-anchored complement to the technology-readiness-level scale, which tracks maturity but not the cost of advancing it (Mankins 2009; Olechowski et al. 2020).

### 6.2 Rival explanations

Three rival explanations could produce a negative slope that is not a true autonomy learning curve. First, general improvement in computing and software tooling lowers all software-development cost over time; the decade fixed effects address this but a within-decade trend could remain. Second, scale effects, where larger and better-funded missions field autonomy more efficiently, could masquerade as heritage; the NICM-class normalization addresses this. Third, selection, where cheap demonstrations are attempted only after heritage exists, could bias the association; the maturation covariate and cross-class robustness specification address this. The discussion of any fitted result will weigh these rivals explicitly.

### 6.3 External validity and the anchor frameworks

Mokyr's framework predicts that the slope will be heterogeneous and steeper where the propositional base is mature and the knowledge is codified and reusable, for example through the Core Flight System substrate (cFS community 2016). This is testable as effect heterogeneity across capability classes and is the most policy-relevant secondary finding the design can produce. Arthur's framework predicts that the realized slope is path-dependent: the classes that received early investment accumulate their own increasing returns, so the measured average slope reflects historical sequence as much as intrinsic learnability (Arthur 1994; Arthur 2021). This caution bounds the external validity of any single estimated learning rate and argues against treating the slope as a universal constant.

### 6.4 What would falsify the contribution

The contribution is falsified by a fitted slope that is zero or positive, or whose confidence interval contains zero across the baseline and robustness specifications. It is also weakened, though not strictly falsified, if the apparent slope collapses once the maturation covariate or the cross-class heritage measure is added, which would indicate that the raw association reflected maturity or shared components rather than within-class flight heritage.

---

## 7. Contribution and Conclusion

This dissertation specifies a falsifiable test of a widely held but unmeasured assumption: that each successive onboard-autonomy flight demonstration lowers the cost to field the next. The contribution is a single measurement, the experience-curve slope of capability-class autonomy qualification cost on cumulative flight-demonstrated heritage, estimated by a Wright/Henderson log-log regression with capability-class and decade fixed effects on a panel built from NASA TechPort records, NTRS autonomy demonstration reports, and NICM-class cost estimates. The hypotheses are stated so that either outcome is informative: a negative, significant slope rejects the flat-cost null and yields a planning parameter; a slope indistinguishable from zero supports the null and reframes the portfolio rationale around capability value. The work is presented honestly at the design stage, with full development of the data construction, identification strategy, and threats to validity, and with expected results clearly labeled as illustrative and not yet estimated. The two anchor frameworks sharpen the test rather than decorate it: Arthur identifies the learning-effect mechanism the curve measures and warns that the realized path is contingent, and Mokyr predicts that the slope will be steeper where the underlying knowledge is mature, codified, and reusable. The completed study will deliver a defensible number, or a defensible failure to find one, on a question that NASA and JPL currently answer by assumption.

---

## References

1. Arthur, W. B. (1989). Competing Technologies, Increasing Returns, and Lock-In by Historical Events. *The Economic Journal*, 99(394), 116-131. doi: [10.2307/2234208](https://doi.org/10.2307/2234208)
2. Arthur, W. B. (1994). *Increasing Returns and Path Dependence in the Economy*. University of Michigan Press. doi: [10.3998/mpub.10029](https://doi.org/10.3998/mpub.10029)
3. Arthur, W. B. (2021). Foundations of complexity economics. *Nature Reviews Physics*, 3, 136-145. doi: [10.1038/s42254-020-00273-3](https://doi.org/10.1038/s42254-020-00273-3)
4. Balaram, J., Aung, M., Golombek, M. P. (2021). The Ingenuity Helicopter on the Perseverance Rover. *Space Science Reviews*, 217, 56. doi: [10.1007/s11214-021-00815-w](https://doi.org/10.1007/s11214-021-00815-w)
5. Bhattacharya, S., et al. (2017). The cost-quantity relations and the diverse patterns of learning by doing: Evidence from India. *Research Policy*, 46(10). doi: [10.1016/j.respol.2017.09.005](https://doi.org/10.1016/j.respol.2017.09.005)
6. Chien, S., et al. (2005a). Using Autonomy Flight Software to Improve Science Return on Earth Observing One. *Journal of Aerospace Computing, Information, and Communication*, 2(4), 196-216. doi: [10.2514/1.12923](https://doi.org/10.2514/1.12923)
7. Chien, S., et al. (2005b). Monitoring active volcanism with the Autonomous Sciencecraft Experiment on EO-1. *Remote Sensing of Environment*, 101(4). doi: [10.1016/j.rse.2005.08.007](https://doi.org/10.1016/j.rse.2005.08.007)
8. Estlin, T. A., et al. (2012). AEGIS Automated Science Targeting for the MER Opportunity Rover. *ACM Transactions on Intelligent Systems and Technology*, 3(3). doi: [10.1145/2168752.2168764](https://doi.org/10.1145/2168752.2168764)
9. Farmer, J. D., Lafond, F. (2016). How predictable is technological progress? *Research Policy*, 45(3), 647-665. doi: [10.1016/j.respol.2015.11.001](https://doi.org/10.1016/j.respol.2015.11.001)
10. Francis, R., et al. (2017). AEGIS autonomous targeting for ChemCam on Mars Science Laboratory: Deployment and results of initial science team use. *Science Robotics*, 2(7). doi: [10.1126/scirobotics.aan4582](https://doi.org/10.1126/scirobotics.aan4582)
11. Gao, Y., Chien, S. (2021). Autonomy for Space Robots: Past, Present, and Future. *Current Robotics Reports*, 2, 251-263. doi: [10.1007/s43154-021-00057-2](https://doi.org/10.1007/s43154-021-00057-2)
12. Lafond, F., et al. (2018). How well do experience curves predict technological progress? A method for making distributional forecasts. *Technological Forecasting and Social Change*, 128, 104-117. doi: [10.1016/j.techfore.2017.11.001](https://doi.org/10.1016/j.techfore.2017.11.001)
13. Mankins, J. C. (2009). Technology readiness assessments: A retrospective. *Acta Astronautica*, 65(9-10), 1216-1223. doi: [10.1016/j.actaastro.2009.03.058](https://doi.org/10.1016/j.actaastro.2009.03.058)
14. Nagy, B., Farmer, J. D., Bui, Q. M., Trancik, J. E. (2013). Statistical Basis for Predicting Technological Progress. *PLOS ONE*, 8(2), e52669. doi: [10.1371/journal.pone.0052669](https://doi.org/10.1371/journal.pone.0052669)
15. Olechowski, A. L., Eppinger, S. D., Joglekar, N., Tomaschek, K. (2020). Technology readiness levels: Shortcomings and improvement opportunities. *Systems Engineering*, 23(4), 395-408. doi: [10.1002/sys.21533](https://doi.org/10.1002/sys.21533)
16. Perez, C. (2009). Technological revolutions and techno-economic paradigms. *Cambridge Journal of Economics*, 34(1), 185-202. doi: [10.1093/cje/bep051](https://doi.org/10.1093/cje/bep051)
17. Stahl, H. P., Henrichs, T., Luedtke, A., West, M. (2010). Single-variable parametric cost models for space telescopes. *Optical Engineering*, 49(7), 073006. doi: [10.1117/1.3456582](https://doi.org/10.1117/1.3456582)
18. Thompson, P. (2012). The Relationship between Unit Cost and Cumulative Quantity and the Evidence for Organizational Learning-by-Doing. *Journal of Economic Perspectives*, 26(3), 203-224. doi: [10.1257/jep.26.3.203](https://doi.org/10.1257/jep.26.3.203)
19. Verma, V., et al. (2023). Autonomous robotics is driving Perseverance rover's progress on Mars. *Science Robotics*, 8(80). doi: [10.1126/scirobotics.adi3099](https://doi.org/10.1126/scirobotics.adi3099)
20. Center for Security and Emerging Technology (2024). Shaping the U.S. Space Launch Market. doi: [10.51593/20240017](https://doi.org/10.51593/20240017)
21. International Astronautical Federation (2024). Impact of Launch Cadence on the Automation and Economics of Constellation Operations. doi: [10.52202/078367-0036](https://doi.org/10.52202/078367-0036)
22. Bernard, D. E., et al. (1999). Validation and Verification of the Remote Agent for Spacecraft Autonomy. NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20210003369](https://ntrs.nasa.gov/citations/20210003369)
23. Smith, B. D., et al. (2000). Challenges and Methods in Testing the Remote Agent Planner. NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20210001679](https://ntrs.nasa.gov/citations/20210001679)
24. Bernard, D. E., et al. (1997). Infusion of Autonomy Technology into Space Missions: DS1 Lessons Learned. NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20210003565](https://ntrs.nasa.gov/citations/20210003565)