Master Thesis Project - 2027

1. Bridging the sim-to-real Gap: Domain randomization and synthetic data for robotic arm control (STHLM) Background & Description We offer a master's thesis project on closing the gap between simulation and physical hardware in robot learning. Policies trained purely on simulated data, including Vision-Language-Action (VLA) models, often fail on real hardware due to mismatches in visuals, physics, and sensing. Domain randomization addresses this by varying simulation parameters during data generation so the model learns features invariant to the sim-real gap rather than simulator artifacts (Tobin et al., 2017). This thesis treats synthetic data production as an optimization problem: which parameter distributions, quantities, and sim/real mixtures maximize real-world performance per unit of data and compute? Recent work shows that even simple sim/real co-training recipes substantially improve manipulation success rates (Maddukuri et al., 2025), but principled, mathematically grounded strategies remain an open question. Core idea: synthetic data is generated via a mathematically justified and data-efficient randomization strategy, a VLA or comparable model is trained on it, and the resulting policy is deployed on physical robotic arm hardware. The theoretical contribution is the mathematical analysis behind the strategy, for example coverage guarantees, sample complexity, or framing parameter selection as an optimization problem. The applied contribution is validating the policy on real hardware. Students will work with a physical robotic arm, GPU compute, and guidance from Modulai's ML engineers Example directions Formal analysis of how domain randomization ranges should be chosen relative to the true (unknown) distribution of real-world conditions, and what guarantees this gives on real-world generalization Optimizing the mixture and scheduling of simulated versus real demonstration data during VLA fine-tuning Automatic or learned domain randomization, where randomization parameters are adapted based on validation performance rather than fixed by hand End-to-end evaluation: train in simulation only, train with a randomization strategy, and train with sim+real co-training, then compare real-arm task success rates ML Techniques and Tools Python, PyTorch, Git, Hugging Face Robotics simulators (e.g. MuJoCo, Isaac Sim, or similar) for synthetic data generation Domain randomization and sim-to-real transfer methods Vision-Language-Action models and other end-to-end control architectures Statistical and optimization methods for data generation strategy design Real-time control and deployment on physical robotic arm hardware References Tobin et al., Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World, 2017. arXiv:1703.06907 - https://arxiv.org/abs/1703.06907 Maddukuri et al., Sim-and-Real Co-Training: A Simple Recipe for Vision-Based Robotic Manipulation, 2025. arXiv:2503.24361 - https://arxiv.org/abs/2503.24361 2. Image representations in property valuation (Computer vision/Tabular) (Sthlm/Gbg) Background & Description We offer a master's thesis project on using image data to improve the accuracy of automated property valuation. The project is run together with a growing startup that is building a state-of-the-art valuation engine, with guidance from Modulai's ML engineers. Automated valuation models traditionally rely on tabular data: living area, number of rooms, location, construction year and historical transactions. Two apartments with near-identical records can still differ substantially in market value because of condition, renovation standard, light, layout and view. Much of that residual signal is present in listing photographs, floor plans and aerial imagery, but is rarely exploited beyond coarse heuristics. Early work showed that a learned "luxury level" derived from interior and exterior photos, combined with metadata, can outperform established metadata-only estimates (Poursaeed et al., 2017), and later studies confirm that visual features add predictive power on top of strong tabular baselines (Kostic & Jevremović, 2021). This thesis treats the image side as a representation and integration problem: which visual representations carry the signal that tabular features miss, and how should they be fused into a production valuation model without hurting robustness, calibration or explainability? The candidate representations span the full toolbox - image classification (room type, condition, renovation standard), semantic segmentation (materials, surfaces, greenery, floor-plan geometry), object detection (fireplaces, appliances, balconies) and general-purpose embeddings from pretrained vision or vision-language backbones. Core idea: Explore image model approaches to represent the image information as efficiently as possible, while keeping explainability of the model. Investigate how such methods may contribute to improved valuation accuracy for apartments and house listings, The methodological contribution is the comparison and fusion strategy; the applied contribution is a validated improvement in a system that is actually shipped. Students will work with large-scale real listing data, GPU compute, and close guidance from both the company's ML team and Modulai's ML engineers. It is also a domain that is unusually easy to relate to - everyone lives somewhere. Example directions Systematic comparison of representation families - classification heads, segmentation masks, detected objects and raw embeddings - on equal footing, measured as marginal accuracy over a strong tabular baseline Off-the-shelf vision-language models used as attribute extractors versus purpose-trained models and learned embeddings: accuracy, cost and latency per valuation Fusion architecture: late fusion of pooled embeddings into a gradient-boosted model versus end-to-end multimodal training, including how to aggregate a variable number of images per property Weak