New benchmark tests robot navigation in crowded spaces with social awareness

Robots that follow natural language instructions while navigating indoors or outdoors must learn more than just reaching a destination—they need to move safely around people. A new unified benchmark called HA-VLN 2.0 addresses this gap by introducing social-awareness requirements into vision-and-language navigation tasks.

The challenge of human-aware navigation#

Vision-and-Language Navigation (VLN) research has traditionally focused on agents finding their way through either discrete grid-like spaces or continuous 3D environments, but typically in isolation. Real-world deployment demands that autonomous systems operate in crowded, dynamic settings where respecting personal space and avoiding collisions with multiple people becomes critical. HA-VLN 2.0 bridges this divide by unifying discrete and continuous navigation scenarios under a single framework that prioritizes human-centric safety.

What the benchmark delivers#

The team released four major contributions to advance the field:

• Standardized task and metrics that measure both navigation accuracy toward a goal and adherence to proxemics—the spaces people expect around them
• HAPS 2.0 dataset and simulators covering 16,844 socially grounded instructions with multi-human interactions, outdoor settings, and tighter alignment between language and motion
• Comprehensive benchmarking showing that state-of-the-art navigation agents experience sharp performance drops when dealing with human dynamics and partial observability (agents can't see around corners)
• Real robot validation demonstrating that strategies learned in simulation transfer effectively to physical systems, backed by an open leaderboard for transparent comparison

Why social modeling matters#

Testing results underscored a critical finding: explicit modeling of social dynamics improves navigation robustness and reduces collisions. When agents were equipped with awareness of human presence and spatial preferences, they navigated more safely than systems optimized purely for task completion. This underscores the necessity of moving beyond goal-centric approaches toward human-centric design in autonomous systems.

By releasing datasets, simulators, baseline implementations, and evaluation protocols, HA-VLN 2.0 establishes a foundation for research into safe, socially intelligent robotic navigation.