392243 Project: Confidence-Aware Policies: Using Uncertainty to Decide When a Policy Should Act (Pj) (SoSe 2026)

Real robots should not act blindly. A practical idea is to estimate how confident a policy (or value function) is, and only let the policy act when confidence is high. In simulation, we can test this idea rigorously and cheaply. In this project, you will implement an uncertainty signal (e.g., by comparing several models or using a model that outputs a distribution of outcomes) and wrap it around an existing learning pipeline. If uncertainty rises above a threshold, the system should trigger a simple fallback behavior. You will then measure how this changes success rates, failure frequency, and how often the fallback is triggered. The emphasis is on building something robust and easy to interpret, with clear plots and a principled evaluation. Solid Python skills and comfort reading/modifying existing ML code. Some familiarity with PyTorch training loops is helpful. You do not need deep theory; the project is primarily about building reliable monitoring and evaluating it carefully.