MERGE: Guided Vision-Language Models for Multi-Actor Event Reasoning and Grounding in Human-Robot Interaction

Abstract

We introduce MERGE, a system for situational grounding of actors, objects, and events in dynamic human-robot group interactions. Effective collaboration in such settings requires consistent situational awareness, built on persistent representations of people and objects and an episodic abstraction of events. MERGE achieves this by uniquely identifying physical instances of actors (humans or robots) and objects and structuring them into actor-action-object relations, ensuring temporal consistency across interactions. Central to MERGE is the integration of Vision-Language Models (VLMs) guided with a perception pipeline: a lightweight streaming module continuously processes visual input to detect changes and selectively invokes the VLM only when necessary. This decoupled design preserves the reasoning power and zero-shot generalization of VLMs while improving efficiency, avoiding both the high monetary cost and the latency of frame-by-frame captioning that leads to fragmented and delayed outputs. To address the absence of suitable benchmarks for multi-actor collaboration, we introduce the GROUND dataset, which offers fine-grained situational annotations of multi-person and human-robot interactions. On this dataset, our approach improves the average grounding score by a factor of 2 compared to the performance of VLM-only baselines - including GPT-4o, GPT-5 and Gemini 2.5 Flash - while also reducing run-time by a factor of 4. The code and data are available at www.github.com/HRI-EU/merge.

Figures (4)

• Figure 1: Overview of MERGE. The Action Reasoner is a VLM that integrates three structured inputs: a region of the current camera frame centered on person_y to capture the relevant scene entities and their spatial arrangement; reference images of previously detected objects_x to ground reasoning in known instances; and the four recent images captured by image_i. These inputs are sourced from the Memory module, which aggregates person locations from the Person Tracker and object references from the Object Detector at system initialization. In parallel, a lightweight Action Detector continuously predicts action labels from the image stream, and any change in a person’s predicted action triggers the Action Reasoner, with the updated label itself serving as an additional VLM input.
• Figure 2: Visualization of the prompt provided to VLM. The prompt begins with a general introduction, followed by cropped object and person images (with person ids), each uniquely identifiable via caption. The robot hand is optionally included to assess interaction. The last four images show the recent captured images before the action trigger. The prompt concludes with a task description guiding the VLM through action inference, object assignment, spatial relation, and robot interaction.
• Figure 3: GROUND-Train provides a rich set of annotations captured from four synchronized camera views. Each video includes person-wise action segmentation labels, 2D pose annotations along with human and object bounding boxes across all views, with the pose and human box annotations further linked through cross-view tracking.
• Figure 4: Left: Example image from GROUND-Eval captured from the robot’s perspective, showing two people and the robot sorting fruits onto two plates. Right: Front-facing view of the robot.