An end-to-end transformer-based model for interactive grounded language understanding

This paper delves into Interactive Grounded Language Understanding (IGLU) problems within the context of Human-Robot Interaction (HRI), where a robot interprets user commands about the environment. In this scenario, the robot's objective is to determine if a given command can be executed within the environment. If ambiguity or incomplete information is detected, the robot generates pertinent clarification questions. Drawing inspiration from the GrUT framework and employing a BART-based model that combines the user's utterance with the description of the environment, this study evaluates the applicability of the GrUT approach in an end-to-end Grounded QG task. The assessment of question quality is conducted through both automated metrics and human evaluation. While the results highlight the proficiency of the BART-based method in question generation, challenges arise due to dataset limitations from the IGLU competition at NeurIPS 2022. Nevertheless, this research provides valuable insights into BART's generative capabilities in the realm of HRI.