Introducing a framework to reduce foundation models hallucinations in robotics application

Damien Alouges, Raphaël Lallement, Matteo Morelli, Sébastien Gérard

Keywords: Foundation models, Hallucination mitigation, Framework, Robotics

TL;DR: Presentation of a formalization of a framework to reduce hallucinations of foundation models, and an experiment to mesure its efficiency on a simple use case.

Abstract:

While Foundation Models (FM) hold potential for enhancing the autonomy and common sense of robotic AI systems, their tendency to hallucinate makes them unreliable in real-world applications. We seek to eliminate such hallucinations to enable controlled and trustworthy outputs for robotic agents.
In the pursuit of this objective we initiated the definition and the implementation of a Critics framework inspired by a positional paper, the framework uses expert-rules-based critics to detect hallucinations and request a new generation with the hallucination pointed out to the foundation model. We propose a formalization of the framework, defining the concepts used for now and ideas for future upgrades.
We did a first implementation on a basic scene-description use case to identify the main challenges of this approach.
The results are promising as we observed a positive impact on the number of hallucinations the framework produced versus what the foundation model alone generated.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 3