HA-VLN 2.0: Open Benchmark for Human-Aware Navigation in Dynamic Environments
HA-VLN 2.0 establishes a comprehensive benchmark for Vision-and-Language Navigation (VLN) by incorporating specific social-awareness constraints to enhance navigation in both discrete and continuous environments. This benchmark features a standardized task along with metrics that evaluate goal accuracy and respect for personal space. It comprises the HAPS 2.0 dataset and simulators that replicate multi-human interactions, outdoor settings, and precise language-motion coordination. Analysis of 16,844 socially grounded instructions indicates significant performance declines for top agents when faced with human dynamics and partial observability. Experiments with real-world robots confirm the effectiveness of sim-to-real transfer, and an open leaderboard facilitates transparent evaluations. Findings highlight that explicit social modeling enhances navigation resilience and minimizes collisions, emphasizing the importance of human-centric strategies. The project is available on arXiv.
Key facts
- HA-VLN 2.0 introduces explicit social-awareness constraints to VLN.
- The benchmark covers discrete and continuous environments with dynamic multi-human interactions.
- HAPS 2.0 dataset and simulators model multi-human interactions and outdoor contexts.
- Benchmarks were run on 16,844 socially grounded instructions.
- Leading agents show sharp performance drops under human dynamics and partial observability.
- Real-world robot experiments validated sim-to-real transfer.
- An open leaderboard enables transparent comparison of methods.
- Explicit social modeling improves navigation robustness and reduces collisions.
Entities
—