[STAGE MASTER 2 ]

Description In industry 5.0 marks a new stage in the evolution of the industrial world, built on three key pillars: human-centricity, sustainability, and resilience. Rather than focusing solely on productivity, it emphasizes creating systems that respect human capabilities, reduce environmental impact, and remain robust in the face of disruptions. Re-centering the human in industrial systems therefore introduces several challenges (Nahavandi, ), particularly the need to design workspaces that are more ergonomic and compatible with human capabilities. The notion of tool affordance (Gibson, ), borrowed from the social sciences, provides a key framework for understanding how operators perceive the objects in their environment and how they interact with them. By analyzing these interactions, it becomes possible to design technologies that are more intuitive, adapted, and genuinely human-centered. To achieve this, an important component is the accurate detection of objects, especially tools used during industrial tasks. Modern approaches rely on deep learning techniques (Trigka & Dritsas, ), which typically require large amounts of annotated data to reach high performance. However, collecting and manually labelling real-world industrial datasets is costly, time-consuming, and often impractical due to production schedules, safety and confidentiality constraints. This lack of real-world data represents a major limitation for training robust detection models in industrial environments. In this context, synthetic data generated through simulation environments offers a promising direction to enrich training datasets. Our previous work on real/synthetic data ratios has shown that combining even a limited amount of real data with larger volumes of synthetic samples can effectively compensate for the scarcity of real-world observations. By producing large, diverse, and perfectly annotated images at low cost, synthetic data can therefore mitigate data limitations and significantly improve model generalization. Domain randomization techniques applied to synthetic data, such as variations in lighting, object poses, camera viewpoints, and background conditions, can further increase the diversity and realism of synthetic images, helping models become more robust to real-world variability. Traditional data augmentation methods (flip, crop, noise, etc.) can also contribute to improved generalization. Additionally, different transfer-learning strategies can help bridge the gap between synthetic and real-world data by leveraging pretrained models and adapting them to industrial scenarios. As a continuation of the existing work on mixed real/synthetic datasets (Ouarab et al., a; Ouarab et al., b), this internship will explore these advanced techniques using multiple dataset configurations (real, synthetic, and augmented). The objective is to systematically study how domain randomization, data augmentation, and transfer-learning strategies impact the performance and robustness of industrial tool-detection models, and to compare all these methods with each other in order to identify which approaches offer the best performance and generalization in real industrial settings. Work program ● Step 1 : Literature Review (Weeks 2–3) Review key concepts related to synthetic data generation in Unity, data augmentation, and transfer-learning strategies. ● Step 2 : Dataset Preparation and Randomization (Weeks 4–6) Prepare and organize the different datasets: real, synthetic, augmented, and mixed. ● Step 3 : Training and Fine-Tuning Strategies (Weeks 7–14) Train deep learning models for tool detection using different dataset configurations (real, augmented, etc.) and evaluate transfer-learning methods. ● Step 4 : Experimental Evaluation and Analysis (Weeks 15–17) Produce quantitative metrics (mAP-50, F1-score, etc.) under different variations of datasets and transfer-learning techniques, and identify which combinations offer the best generalization to real industrial environments. ● Step 5 : Reporting and Final Presentation (Weeks 18–20) Prepare documentation, dataset summaries, a technical report, and a final oral presentation consolidating the methodology, experiments, and key insights. Bibliography Shorten, C., & Khoshgoftaar, T. M. . A survey on Image Data Augmentation for Deep Learning. Journal of Big Data, 6, 60.Man, M., Arabnia, H. R., & Rasheed, K. . A Review of Deep Transfer Learning and Recent Advancements. Technologies, 11, 40.Nahavandi, S. . Industry 5.0—A Human-Centric Solution. Sustainability, 11, Ouarab, S., Garcia, D., Ragot, N., & Dupuis, Y. . Improving Image-Based Tool Detection in Industrial Workstations using Data Augmentation. In Proceedings of the 51st Annual Conference of the IEEE Industrial Electronics Society (IECON ). Ouarab, S., Garcia, D., Ragot, N., & Dupuis, Y. . Contribution à la caractérisation de l’affordance d’un environnement de travail industriel : une approche basée sur l’apprentissage profond combinant données réelles et synthétiques. In Conférence Nationale sur les Applications de l’Intelligence Artificielle (APIA ), Dijon, France. HAL Id : hal-. Ouarab, S., Boutteau, R., Romeo, K., Lecomte, C., Laignel, A., Ragot, N., & Duval, F. . Industrial Object Detection: Leveraging Synthetic Data for Training Deep Learning Models. In Proceedings, Springer Nature Switzerland, pp. –. Trigka, M., & Dritsas, E. . A Comprehensive Survey of Machine Learning Techniques and Models for Object Detection. Sensors, 25, .Gibson, J. J. . The Ecological Approach to Visual Perception. Houghton Mifflin. Profile Required Profile ● Student in the final year of a Master’s program or engineering school, specializing in computer science, computer vision, artificial intelligence, industrial engineering, or a related field. ● Knowledge of Python programming, basic image processing, and fundamentals of machine learning. ● Experience with the Unity environment is a plus. ● Ability to work autonomously and rigorously, while also collaborating effectively within a multidisciplinary research team. ● Good written and oral communication skills, especially for scientific writing and presenting research results. Starting date Dès que possible