392235 Project: Benchmarking Diffusion and Flow Policies for Reinforcement Learning with a 3D-printed Robotic Arm (Pj) (WiSe 2025/2026)

Recent advances in diffusion and flow-based policies have shown great promise in robotic RL, yet a systematic comparison of these methods on practical robotic manipulation tasks remains missing. This project will benchmark three state-of-the-art algorithms—Diffusion-QL, Flow Q-Learning, and Consistency-AC—in two robotic scenarios (a simple reach task and a more challenging drawer-closing task) using the WidowX robot arm in simulation. Your goal is to measure and compare convergence rates, success metrics, computational efficiency, and stability.

This project offers excellent experience for students interested in cutting-edge algorithmic research, deep learning, and robot manipulation. You will acquire skills in reproducible experimentation and robust algorithmic evaluation. When applicable, your results can be published as a short benchmark note or as an appendix to an existing paper.

For more details or to apply, feel free to contact me directly via email or in-person.

Teilnahmevoraussetzungen, notwendige Vorkenntnisse

You should be proficient in PyTorch and have a solid understanding of deep RL fundamentals and diffusion or flow-based generative models. The project provides you with a prebuilt 3D printed WidowX arm, ready-to-use MuJoCo simulation environments, baseline RL implementations, and all necessary computational resources (though bringing your own GPU is a plus).