Academic Fraud Investigation Report: Generative Adversarial Self-Imitation Learning with Large Language Model Feedback for Robot Control and Navigation (DOI: 10.1109/TRO.2026.3710412)

Academic fraud report · Geng Detector

Summary

This investigation report alleges confirmed academic misconduct in the paper 'Generative Adversarial Self-Imitation Learning with Large Language Model Feedback for Robot Control and Navigation' published in IEEE Transactions on Robotics (DOI: 10.1109/TRO.2026.3710412). The report reaches a verdict of 'confirmed fraud' (实锤). Three main concerns are identified: (1) image reuse and inconsistency in Figure 8, where the backgrounds of Environments 1 and 3 appear identical despite text claims that Environment 1 is an indoor home setting with three pedestrians while Environment 3 contains nine pedestrians in a dense crowd scenario; (2) suspicious reinforcement learning learning curves in Figure 4(a) showing unnatural zigzag confidence interval shading and mirror-symmetric patterns inconsistent with genuine stochastic training; (3) unusually perfect violin plot distributions in Figure 7 that lack the noise and asymmetry typical of real experimental data. The findings suggest potential image fabrication, data manipulation, and visualization beautification. Confidence is high for image reuse and curve irregularities but moderate for statistical visualization claims. Final determination requires official institutional investigation.

Verdict

🔴 Confirmed fraud (实锤) — multiple serious concerns identified regarding image manipulation, suspicious experimental data patterns, and unusual visualization aesthetics.

Key findings

Image reuse with text-figure inconsistency (Figure 8, Page 12): Background images of Environment 1 and Environment 3 are visually identical (matching floor texture, wall color, furniture layout, and lighting), contradicting the paper's text (Page 11) which describes Environment 1 as an indoor home with three pedestrians and Environment 3 as a dense pedestrian scene with nine pedestrians.
Anomalous reinforcement learning learning curves (Figure 4(a), Page 10): Confidence interval shading for GASL³MF and GAL³MF curves shows unnatural regular zigzag oscillations and mirror-symmetric patterns inconsistent with genuine multi-seed RL experimental variance.
Suspiciously perfect violin plot distributions (Figure 7, Page 12): Score distributions (notably GASIL in Environment 2) exhibit overly symmetric, geometrically smooth shapes lacking the noise, skewness, and outliers typical of real robot navigation reward data.

Evidence highlights

Figure 8 visual inspection confirms Environment 1(a) and Environment 3(c) share identical background imagery despite described environmental differences (3 pedestrians in home vs. 9 pedestrians in dense crowd).
Figure 4(a) confidence bands display sawtooth patterns and symmetrical boundaries at multiple time points—behavior not expected from stochastic policy gradient training across random seeds.
Figure 7 violin plots show unnaturally smooth contours and perfect symmetry, suggesting potential synthetic data generation or post-hoc visualization smoothing.
Authors listed: Ke Zhang, Zheng Fang, Enqi Zhao, Zicheng Sun, Jianwu Fang, Jie Huang, Eric Nichols, Randy Gomez, Bo He, Jianru Xue, Guangliang Li.
Journal: IEEE Transactions on Robotics, Year: 2026.
DOI preserved: 10.1109/TRO.2026.3710412.

Notes

All three findings are marked as '✅ Verified' in the original report.
Findings 1 and 2 are rated severity 🔴 (severe); Finding 3 is rated 🟠 (moderate).
Recommendations include contacting authors for raw training logs, posting concerns on PubPeer, and filing complaints with IEEE Transactions on Robotics editorial board and authors' institutions (Ocean University of China, Xi'an Jiaotong University, etc.).
This is an AI-assisted preliminary analysis; official institutional investigation is required for definitive determination of academic misconduct.
Uncertainty remains regarding whether anomalous patterns stem from fabrication, inappropriate smoothing, or rendering artifacts; raw data examination would be necessary to confirm.