I built a complete teleoperation setup to control a robot hand and arm, using this to collect demonstrations and train an adapted 3D Diffusion Policy. My focus was on specific industrial scenarios involving the handling of multiple identical parts. This is normally a nightmare for visuomotor policies because the model cannot distinguish between identical objects based solely on the global features of the vision encoder. Current solutions relying on textual conditioning are insufficient for the industrial context.In addition to conducting experiments on the behaviour of the policy with unknown objects and areas, I also found a method to train the model to indicate when it has finished the task so it can reliably switch back to a classical motion planner.

Developed as part of a university project at Heilbronn University, this real-time traffic sign recognition system uses YOLOv8n for fast and robust detection. It handles challenging conditions like occlusion, poor lighting, and complex backgrounds by leveraging a custom synthetic dataset, multi-stage classification, and real-time frame filtering.

Stabled Grounding SAM is a powerful tool for generating synthetic datasets with pre-segmented images. It combines Stable Diffusion, Grounding DINO, and Segment Anything to create annotated datasets from just a single input image and a label file.