Beste Aydemir

I am an MSc student in Data Science at Ludwig-Maximilian University of Munich.

My research interests are vision language models and their applications to embodied agents and autonomous driving.

Email  /  GitHub  /  LinkedIn

• Relevant Courses: Statistical Reasoning and Inference, Predictive Modeling, Advanced Deep Learning for Robotics

• Relevant Courses: Probability and Statistics, Fundamental Structures of Computer Science, Neural Networks, Algorithms and Programming I-II, Stochastic Models
• Signals and Systems, Linear Algebra and Differential Equations, Nonlinear Systems Theory

• Addressing the problem that in streaming multimodal data, questions and their answers can appear at different times
• Introducing a three-module architecture: memory retention, evidence retrieval, and evidence-grounded trigger for answering at the right time
• Including AnytimeVQA-1K, a new benchmark of 1,000 Q&A pairs over 189 long videos to test temporal awareness

• Class-agnostic 3D instance masks generated using Mask3D + OpenMask3D pipelines, which use CLIP features and 2D SAM masks on SceneFun3D dataset
• Building hierarchical open-vocabulary scene graph with CLIP embeddings, and using Qwen3-14B to extract contextual objects, spatial relations, and functional components
• Using PartField (trained on Objaverse) to decompose point cloud into functional parts (e.g., handles, knobs, cushions).
• Qwen2.5-VL-3B selects the correct functional part from segmented components using labeled 2D renderings

• Utilized multimodal models such as InstructBLIP and Video-Llava for in-context learning on driving datasets
• Implemented text based memory systems for LLM agents in streaming environments

• Replication of "Planning with Diffusion for Flexible Behavior Synthesis" (Janner et al., 2022) in for 2D Maze environment in PyTorch
• Extension of the paper to 3D environment of robotic arm with 7 DoF
• Training on Google Cloud Engine

• Using video language models for autonomous driving tasks
• Experiments on fine-grained action understanding of videos in traffic scenarios
• Utilizing existing autonomous driving datasets to curate a more descriptive dataset to be used for retrieval

• Simulation of centralized MPC control of multi-agent drone system for navigations, and experiments
• Developing learning-based alternatives to bypass linear solvers of CasADi for speedup

• Q-Learning based models for behavioral analysis in the Morris Water Maze experiment
• Developed the cognitive value matric to assess the spatial navigation capabilities in the simulations
• Featured a poster at the 49th European Brain and Behaviour Society (EBBS) Meeting

Design and source code from Jon Barron's website