AI Drone Flight-Time Study: 26.7% Endurance Claim and R²=0.97 Audited

What is the 2026 AI drone endurance study?

The Atlantis Press proceedings chapter, published 31 March 2026, by Durishetti Sathish Kumar and co-authors, is titled Performance Analysis of AI Tools for Drone Endurance Optimization (ICAISDA 2025 proceedings, pp. 474–490; DOI 10.2991/978-94-6239-616-6_36). It describes Random Forest, XGBoost, genetic algorithms and reinforcement learning for predicting or optimising drone endurance. The publisher's abstract reports 26.7% greater flight time, an R² of 0.97 for reinforcement learning, and claimed energy and discharge-related improvements.

Publishing these results is not equivalent to independently reproducing them. UXV Searcher has independently recalculated percentages from the paper's stated numbers, checked its dataset citation, and examined methodological gaps, but has not tested the aircraft or obtained the underlying prediction and power measurements. Numerical accuracy and experimental reproducibility must remain separate assessment categories.

Is the reported 26.7% drone flight-time increase correct?

The original paper, Section 4.1 and Figure 7 provides four payload comparisons. We use (optimised − baseline) ÷ baseline × 100 for each reported pair.

The largest reported percentage is (19 − 15) ÷ 15 × 100 = 26.7% at 2.0 kg. This is four additional reported minutes in one scenario, not 26.7% more battery capacity and not a general result for all aircraft. The publication does not provide a clear independent replication sample, confidence interval or complete flight-test setup supporting those four pairs.

Payload and speed matter: Table 1 is a different set of six scenarios

The paper's Table 1 lists the following results by payload and horizontal speed. We independently recomputed each percentage from the printed duration pairs.

Unresolved scenario reconciliation: Figure 7 gives 0.5 kg at 22 → 26 minutes, while Table 1's 0.5 kg / 10 m/s row gives 22 → 27. Figure 7 gives 1.0 kg at 20 → 24, while Table 1's 1.0 kg / 10 m/s row gives 20 → 25. The authors do not disclose Figure 7 speed and scenario controls sufficient to establish whether these are genuinely different test setups or inconsistent reporting. The shared 2.0 kg 15 → 19 case does not prove the other comparisons can be pooled. Do not combine these rows as independent repeat flight measurements or derive an average performance guarantee.

What does R² = 0.97 mean for the reinforcement-learning model?

Table 2 of the research chapter reports these prediction-error and goodness-of-fit metrics:

The reported R² = 0.97 describes the fit of the RL model's endurance predictions relative to target variability in the authors' evaluation; it does not mean a 97% mission-completion rate, 97% safe flight, 97% model certification or guaranteed endurance. The RL-agent MAE of 0.8 minutes and RMSE of 1.0 minute are also paper-reported prediction errors, not independently instrumented flight precision.

R² cannot be independently recomputed using only this four-row summary. An external reviewer would need actual observed flight durations, each predicted duration, how the data were partitioned, model settings and the test protocol. The article does not include that prediction-level evidence or a reconciled source dataset. Therefore, R² is independently verified as a reported number, not as a reproduced statistical result.

Did the paper demonstrate 13.3% greater drone battery capacity?

No. The paper discusses a decrease in maximum battery discharge percentage, not a verified increase in physical cell energy storage. Figure 8 and its surrounding text report the following payload comparisons:

The 0.5 kg case changes 90% → 78%, a 12-percentage-point decrease or 13.3% relative reduction. At 2.0 kg, 98% → 85% represents 13 percentage points or a 13.3% relative decrease. These figures do not establish that the battery has more rated watt-hours or an improved cycle-life capacity. The paper uses imprecise wording when describing the reduction as an “increase” in battery consumption or utilisation. Percentage points and percentage changes should be stated separately.

The abstract also claims up to 22.5% reduction in energy per kilometre (Wh/km). We could not independently recompute that figure from traceable route-specific energy/distance observations in the published tables. It should be treated as an unreconciled author-reported abstract claim rather than a validated benchmark for OEM or battery procurement.

Critical methodology gap: an image-detection dataset is not endurance telemetry

Section 3.1 identifies the Drone Dataset (UAV), referenced as item 19 and describes 1,359 labelled drone images and 1,486 drone objects. The cited Kaggle material concerns visual drone object detection, not a documented collection of measured flight minutes, usable battery energy, power draw, route length, wind, payload and airspeed. The paper does not clearly explain how object-detection image labels were converted into the training or validation observations for the endurance prediction models.

The article describes 1,012 training images and 347 validation images, exactly 1,359 in total. Recalculation gives approximately 74.5% training and 25.5% validation. Elsewhere the methodology states a 70% training / 15% validation / 15% testing split. Those allocations do not match, and the stated 1,012/347 allocation does not identify a separate held-out test subset. Without the actual data and partitions, claimed independent model validation cannot be reproduced.

The paper also does not provide a complete reconciled aircraft and battery specification, raw per-flight measurements, prediction-target pairs or uncertainty intervals. The discrepancy does not automatically mean every printed value is wrong; it means the research is not sufficiently reproducible to support a product-specific commercial endurance guarantee.

What should UAV battery suppliers and procurement teams validate next?

Before using a reported AI endurance benefit in acquisition, technology comparison or market forecasts, buyers need the exact aircraft type and software build, baseline and optimised mission profiles, battery chemistry and state of health, usable watt-hours, payload, airspeed, environmental conditions, controlled repeat flight logs, measured power per distance, model features and prediction-level results. Require sample sizes, variability and clearly partitioned training and held-out evaluation data.

In UXV Searcher's research taxonomy, distinguish source-reported numerical example, independently recalculated percentage, internally consistent study design, replicated physical flight test and customer-accepted product performance. This paper currently supports the first two classifications, not the latter three. The US Department of Energy's 2022 AI and UAS risk presentation and the NIST AI Risk Management Framework offer useful general evidence-governance context, but they are not separate validation of the paper's model or battery experiments.

For commercial sizing, do not multiply the study's 26.7% figure across an industry fleet or conflate model-driven flight planning with nominal battery capacity. Software licensing, onboard compute, physical batteries, flight testing and integration services should be modelled separately, with duplicate hardware revenue removed. Where performance validation, customer adoption or annual units remain undisclosed, mark them NOT ESTABLISHED.