Coherence-Driven Multimodal Safety Dialogue with Active Learning for Embodied Agents

TL;DR

This work tackles safety-aware interaction for embodied agents by introducing M-CoDAL, a multimodal dialogue system that leverages coherence relations (PDTB and SDRT) to improve contextual understanding. It combines clustering-based active learning with an external LLM to identify informative safety-instance transcripts, and uses distillation from a high-capacity model to train a lightweight learner. A novel 1K safety-violation dataset derived from 2K Reddit images is created, annotated, and used to train and evaluate the system, with automated metrics indicating improved safety resolution and sentiment, and a real-world deployment on a Hello Robot Stretch showing heightened persuasiveness. The results demonstrate transferability of the active-learning benefits to models not included in the loop, and real-user study confirms the system’s practical impact in persuasive, safety-focused embodied dialogue, while also highlighting user comfort considerations for future personalization.

Abstract

When assisting people in daily tasks, robots need to accurately interpret visual cues and respond effectively in diverse safety-critical situations, such as sharp objects on the floor. In this context, we present M-CoDAL, a multimodal-dialogue system specifically designed for embodied agents to better understand and communicate in safety-critical situations. The system leverages discourse coherence relations to enhance its contextual understanding and communication abilities. To train this system, we introduce a novel clustering-based active learning mechanism that utilizes an external Large Language Model (LLM) to identify informative instances. Our approach is evaluated using a newly created multimodal dataset comprising 1K safety violations extracted from 2K Reddit images. These violations are annotated using a Large Multimodal Model (LMM) and verified by human annotators. Results with this dataset demonstrate that our approach improves resolution of safety situations, user sentiment, as well as safety of the conversation. Next, we deploy our dialogue system on a Hello Robot Stretch robot and conduct a within-subject user study with real-world participants. In the study, participants role-play two safety scenarios with different levels of severity with the robot and receive interventions from our model and a baseline system powered by OpenAI's ChatGPT. The study results corroborate and extend the findings from the automated evaluation, showing that our proposed system is more persuasive in a real-world embodied agent setting.

Figures (5)

• Figure 1: Our proposed dialogue system first parses safety violation in an image with Penn Discourse Treebank Prasad2008ThePD relations, and then generate an appropriate response by choosing an Segmented Discourse Representation Theory relation Asher2005LogicsOC.
• Figure 2: In our active learning loop, informative instances are identified by an external LLM using a composite score. These are distilled using another LLM for coherence relations and responses and are used to retrain the learner LLM.
• Figure 3: Distribution of safety keywords in our dataset. Safety violations with mold and stove are more prevalent.
• Figure 4: Examples of images retrieved from Reddit and their safety violations, obtained through LMM and manual correction.
• Figure 5: Setup for user study. A Hello Robot Stretch robot observes the surroundings, identifies a safety violation in the scene, and engages in conversation with a user.