392242 Project: Learning Rewards from Videos: Evaluating Pretrained Reward Models and Using Them for Robot Policy Learning (Pj) (SoSe 2026)

Pretrained "reward models" (PRM) promise to estimate task progress directly from observations (e.g., videos), which could reduce the need for hand-designed rewards in reinforcement learning (RL). In this project, you will evaluate an existing PRM on a simulated robotic manipulation task: does it track progress and predict success zero-shot, and where does it fail? If the signal is usable, you will design a simple policy improvement experiment that uses the reward model as-is (zero-shot), for example by ranking, filtering, or weighting trajectories, and compare against naive supervised baselines. The project is simulation-only and focuses on careful evaluation and clean experimentation. You will be given a mature codebase for the simulation, training loops and utilities. Requirements: Solid Python skills and comfort working with existing ML code and datasets. Prior RL / Imitation Learning experience is helpful but not required.