Research map / Vision-language and multimodal / Robot action and navigation
Human-Aware Vision-and-Language Navigation: Bridging Simulation to Reality with Dynamic Human Interactions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces HA-VLN, a human-aware vision-and-language navigation task with the HA3D simulator (Matterport3D plus SMPL human motions), the HA-R2R dataset, and two agents (expert-supervised VLN-CM and random-trajectory decision transformer VLN-DT).
Problem
Existing VLN benchmarks assume static environments, panoramic action spaces, and optimal expert supervision, which overstates real-world (Sim2Real) navigation ability.
Approach
Relax three VLN assumptions at once by adding dynamic SMPL-based human activities to Matterport3D, using a 60-degree egocentric action space, and training with sub-optimal or random-walk data, with human-aware metrics (TCR, CR, SR) for evaluation.
Key results
- Best retrained SOTA agent reaches only about 0.40 SR in unseen human-populated environments, versus 0.89 for the Oracle.
- VLN-DT trained on 100% random-walk data achieves 0.21 SR (seen) and 0.11 SR (unseen), comparable to expert-supervised VLN-CM at 0.53 SR under its best setting.
- Zero-shot Airbert navigation error rises by over 167% and SR falls by about 67% when human activities are added.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace panoramic action space with egocentric 60° field of view (ablation) | HA-VLN agents · HA-VLN Validation Seen and Unseen | SR: 0.34 -> 0.20; 0.23 -> 0.16 | Hurt |
| Add dynamic human motion to the environment (ablation) | HA-VLN agents · HA-VLN Validation Seen and Unseen | SR: 0.75 -> 0.40; 0.62 -> 0.50 | Hurt |
| Use sub-optimal instead of optimal expert supervision (ablation) | HA-VLN agents · HA-VLN Validation Seen and Unseen | SR: 0.53 -> 0.50; 0.41 -> 0.40 | Mixed |
| Use 100% random-walk data to train VLN-CM instead of expert data (ablation) | VLN-CM · HA-VLN Validation Seen and Unseen | SR: 0.19 -> 0.03; 0.12 -> 0.02 | Hurt |
| Train on random-walk trajectories without expert supervision (ablation) | VLN-DT · HA-VLN Validation Seen | SR: 0.03 -> 0.21 | Helped |
| Add a penalty-based distance reward (ablation) | VLN-DT · HA-VLN | SR | Helped |
| Add additional penalties for human collisions (ablation) | VLN-DT · HA-VLN | navigation performance | No clear effect |
| Add human activity and dynamic-scene constraints | Airbert · HA-VLN Validation Unseen | SR: 0.75 -> 0.25 | Hurt |
Surprising findings
- Adding sub-optimal expert supervision yields a small SR drop but is claimed to add realism, and additional human-collision penalties in VLN-DT did not significantly improve performance.
- VLN-CM SR declines sharply as the proportion of random-walk data increases, so the expert-supervised agent does not benefit from random data.
Limitations
- Human behavior modeling does not capture social nuances or complex group dynamics.
- HA3D and HA-R2R are restricted to indoor environments.
- Real-world robot evaluation is limited to qualitative office trials that still show collisions with unpredictable humans.
Experimental setup
Datasets: Matterport3D Room-to-Room (R2R) HAPS HA-R2R
Benchmarks: HA-R2R Room-to-Room (R2R)
Metrics: NE SR TCR CR
Baselines: Speaker-Follower Rec (PREVALENT) Rec (OSCAR) Airbert Oracle
Builds on: Matterport3D simulator Room-to-Room Decision Transformer Motion Diffusion Model Speaker-Follower
Closest papers by meaning
- AwareVLN: Reasoning with Self-awareness for Vision-Language Navigation (CVPR 2026 · similarity 0.91)
- D3D-VLP: Dynamic 3D Vision-Language-Planning Model for Embodied Grounding and Navigation (CVPR 2026 · similarity 0.89)
- NavBench: Probing Multimodal Large Language Models for Embodied Navigation (NeurIPS 2025 · similarity 0.88)
- 3D-VLA: A 3D Vision-Language-Action Generative World Model (ICML 2024 · similarity 0.88)
- General Scene Adaptation for Vision-and-Language Navigation (ICLR 2025 · similarity 0.88)
- Dynam3D: Dynamic Layered 3D Tokens Empower VLM for Vision-and-Language Navigation (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.