Research map / Multi-agent systems and games / Human-AI and goal-conditioned
V-IRL: Grounding Virtual Intelligence in Real Life
Jihan Yang, Runyu Ding, Ellis L Brown, Xiaojuan Qi, Saining Xie
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add active detection (adjust ego-pose and focal length) to the urban robot (ablation) | RX-399 open-world detection · Hong Kong street views | accuracy: 0.10 -> 0.63 | Improved |
| add active detection (adjust ego-pose and focal length) to the urban robot (ablation) | RX-399 open-world detection · New York City street views | accuracy: 0.30 -> 0.71 | Improved |
| reclassify GLIP proposals with CLIP instead of using end-to-end open-world detectors | Place detection · V-IRL Place Detection | AR20: 9.0 -> 23.7 | Improved |
| scale CLIP image resolution and model size (ablation) | CLIP place recognition · V-IRL Place Recognition | mAcc (%): 18.2 -> 41.3 | Improved |
| replace oracle landmark information with vision-model recognition (ablation) | VLN agent with GPT-4 · V-IRL VLN-mini | success rate: 1.0 -> 0.44 | Hurt |
| use larger CLIP variant for landmark recognition (B/32 to L/14@336px) (ablation) | VLN agent · V-IRL VLN-mini | success rate: 0.22 -> 0.44 | Improved |
| use LLaVA-1.5-13B instead of CLIP for landmark recognition (ablation) | VLN agent · V-IRL VLN-mini | success rate: 0.44 -> 0.11 | Hurt |
| give the concierge agent real travel times from the platform instead of an ungrounded LLM (ablation) | Diego concierge · NYC itinerary (qualitative) | itinerary feasibility: 30 minutes -> 60--100 minutes | Improved |
| scale up LLaVA generation: LLaVA-NeXT versus LLaVA-1.5 and LLaVA | MLLMs · V-IRL Place VQA | mAcc (%): 61.9 -> 65.9 | Improved |
Closest papers by meaning
- VirtualEnv: A Platform for Embodied AI Research (AAAI 2026 · similarity 0.88)
- Embodied Web Agents: Bridging Physical-Digital Realms for Integrated Agent Intelligence (NeurIPS 2025 · similarity 0.86)
- VS-Bench: Evaluating VLMs for Strategic Abilities in Multi-Agent Environments (CVPR 2026 · similarity 0.85)
- OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments (NeurIPS 2024 · similarity 0.85)
- VitaBench: Benchmarking LLM Agents with Versatile Interactive Tasks in Real-world Applications (ICLR 2026 · similarity 0.85)
- Playing ZendoWorld: Challenging AI Agents on Active Visual Concept Induction (NeurIPS 2026 · similarity 0.84)
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.
