[GB] Evaluating Locally Deployed Large Language Models for Action Planning in Robotics

The integration of Large Language Models (LLMs) into robotic systems opens up new possibilities for automated action planning. The generation of PDDL-like action plans promises a flexible alternative to classical symbolic planners, but also poses challenges with regard to correctness and resource consumption, particularly on embedded systems such as the NVIDIA Jetson Orin. The aim of this thesis is the systematic evaluation of locally executable LLMs of different sizes on a Jetson Orin with regard to their ability to generate correct action plans in a kitchen domain. To this end, an action library, a failure taxonomy, and various prompting techniques are to be developed and integrated into a robotic system via ROS.

Betreuer: Verena Well