ExploraQA: Embodied Question Answering with Long-horizon Proactive Exploration

ICLR 2026 Conference Submission7693 Authors

Keywords: vision language navigation, computer vision

Abstract: Embodied Question Answering (EQA) is a critical task for developing embodied intelligence, requiring agents to autonomously explore environments and answer human questions through perception, navigation, and reasoning. However, existing EQA benchmarks suffer from three key limitations: constrained exploration scope, passive trajectory, and insufficient viewpoint annotation. To address these challenges, we introduce ExploraQA, a large-scale dataset featuring 12,436 diverse, open-ended questions across seven categories, designed to evaluate language, visual, and spatial reasoning. ExploraQA emphasizes long-horizon exploration, proactive trajectory, and comprehensive viewpoint annotations, enabling rigorous assessment of autonomous agents. We further propose an Iterative EQA Data Generation Framework to efficiently produce high-quality annotations via VLMs and human verification. To enhance exploration, we present the Answer Quality-Guided Navigator, which leverages a Topology-Aware Keyframe Search Module for efficient long-range navigation and an Answer Quality Reward Mechanism to optimize question-driven trajectories through dual LLM evaluators. Experimental results show that AQ-Nav achieves a 5.4% absolute improvement in E_score on the ExploraQA unseen test set over state-of-the-art navigators. We will release our dataset and code.

Primary Area: datasets and benchmarks

Submission Number: 7693