Title

COMPASS: Benchmarking safety-aware
navigation for vision language models

Creator

Dolgopolyi, Roman

Abstract

This work presents COMPASS (Compliance-Oriented Model Performance Assessment for Safety Standards), the first open-source bench-mark that unifies spatial navigation, ISO-grounded safety compliance, and instruction following into a single evaluation framework for vision language models (VLMs). COMPASS poses multi-step urban navigation tasks in environments onstructed from real-world Google Street View imagery while enforcing safety constraints derived from ISO 13482 and ISO/TR 4448, operationalized as formal rules that enable measurable
and reproducible evaluation. The benchmark tasks are generated proce-
durally, enabling unlimited data suitable for reinforcement learning with
verifiable rewards (RLVR).

Keywords

artificial intelligence

safety

cybersecurity

Date Available

May 28, 2026

Type

Oral presentation

Course

ITC:4949 - Cybesecurity Capstone project

Session

Oral presentation

Location

Art Center Auditorium

Event

Year 2026 | 10th Annual Student Research and Creative Arts Symposium