Frank Joublin, Antonello Ceravola, Pavel Smirnov, Felix Ocker, Jörg Deigmöller, Anna Belardinelli, Chao Wang, Daniel Tanneberg, Stephan Hasler, Michael Gienger, "CoPAL: Corrective Planning of Robot Actions with Large Language Models", International Conference on Robotics and Automation (ICRA), 2024.

Recent advances in the field of pretrained Large Language Models (LLM) made commonsense knowledge available "out of the box" for a vast range of scenarios including content generation, customer service, and voice assistants. The release of GPT-3.5 (known as ChatGPT) opened prospectives for building highly contextualizable conversational agents, capable to hold a dialog and reflect about various situations as well as on behalf of different social roles (e.g., kitchen chef, software developer). Such an advancement is highly relevant for the robotics domain, where embodied agents have to take care about generating their cyber-physical behaviour in accordance to users’ requests. Robotics behaviour planning comprizes combinations of manipulation and motion planning problems. The contributions of this paper are the following: • Adaptive manipulation planning mechanism with replanning based a pre-simulated feedbacks of different types. • Experimental evaluation of efficiency of the proposed replanning mechanism in two real-world use-cases (pizza, barman) • Evaluation of Improvization capabilities in handling non-standard situations.