392231 Project: Embodied Commonsense Knowledge in Robotic Simulators (Pj) (SoSe 2025)

Intelligent agents are challenged by unknown situations in open worlds. They cannot perform everyday tasks like cutting food or pouring drinks without encountering unknown motions, objects or environments. To mitigate this problem, providing these robots with commonsense knowledge is a possible way to increase their world understanding and support their planning capabilities. However, this household-specific commonsense knowledge is, despite its potential, not often benchmarked or compared between different approaches.

In previous work, we created a language focused, multi-task benchmark for evaluating embodied commonsense capabilities in embodied agents. In this project, the created benchmark should be embedded into an existing robotic / household simulation environment. The students should start with searching for a suitable simulation environment and implement at least two of the five tasks in the benchmark in their chosen environment.

The project can be taken in English or German and, due to its scope, a group of 2-5 people is recommended. However, if no other option is available, individual work is also possible.