HCSG: Human-Centric Semantic-Geometric Reasoning for Vision-Language Navigation
HCSG: Human-Centric Semantic-Geometric Reasoning for Vision-Language Navigation
Authors: Haoxuan Xu, Tianfu Li, Wenbo Chen, Yi ran Liu, Jin Wu, Huashuo Lei, Yunfan Lou, Lujia Wang, Hesheng Wang, Haoang Li Date: 2026-05-13 Paper ID: openalex:2605.13321
Summary
The paper introduces HCSG, a human-centric vision-language navigation (VLN) framework designed to handle dynamic indoor environments with pedestrians. Unlike existing methods that treat humans merely as visual obstacles, HCSG employs a unified Human Understanding Module to perform explicit geometric trajectory forecasting and semantic intent interpretation using Vision-Language Models. These insights, integrated with a social distance loss, enable the navigation agent to perform instruction-conditioned planning while adhering to social norms. Experimental results on the HA-VLNCE benchmark confirm that this active reasoning approach significantly improves both navigational success and safety.
Key Contributions
- Proposes HCSG, the first VLN framework that actively models human intentions and motion rather than treating humans as passive obstacles.
- Introduces a unified Human Understanding Module that synthesizes geometric human pose/trajectory forecasting with semantic action/intent interpretation via VLMs.
- Demonstrates a 14% improvement in Success Rate and a 34% reduction in Collision Rate on the HA-VLNCE benchmark compared to state-of-the-art methods.
Open Questions & Future Work
Key Concepts
- hcsg-framework: A human-centric vision-language navigation framework that integrates semantic intent interpretation and geometric motion forecasting for social navigation.
Archivist Review
I have approved the HCSG framework concept due to its novel combination of VLM-driven semantic intent interpretation with traditional geometric trajectory forecasting in a navigation context. I also approved the open question regarding continuous streaming human-aware navigation, as it highlights a significant limitation in episodic RL/VLN architectures. The second open question was rejected as it is a standard research direction in social robotics that is already well-covered in existing literature.
Approved Concepts
- HCSG Framework: Introduces a novel paradigm for VLN by combining geometric trajectory prediction with semantic action interpretation via VLMs.
Approved Open Questions
- Continuous Human-Aware Navigation Streaming: Transitioning to a continuous streaming setting is essential for practical deployment in realistic human-populated settings, where stop-and-wait behaviors are often impractical or socially disruptive.
Rejected Candidates
- [open_question] Adaptive Context-Aware Social Priors (
adaptive-context-aware-social-priors) - duplicate_existing: The question of adaptive social priors is a common variation of general robotic social navigation constraints and is not sufficiently distinct from existing social-navigation research concerns.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.13321
- paper_id
- 2605.13321
- paper_source
- openalex
- domain
- nlp
- tags
- vision-language-navigationroboticssocial-navigationhuman-intent-predictiondynamic-environments
- architectures
-
- datasets
-
- concept_slugs
- hcsg-framework
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-16T06:03:25Z