[CVPR 2026] SocialNav: Training Human-Inspired Foundation Model for Socially-Aware Embodied Navigation

🎉🎉CVPR 2026 Oral🎉🎉

---

This is the official repository for SocialNav, a foundational model for socially-aware embodied navigation with a hierarchical brain–action architecture. SocialNav unifies high-level social norm understanding with low-level, socially compliant trajectory generation.

📢 Note: This repository contains the model implementation, the training entrypoints and the CityWalker benchmark evaluation script. Pre-trained SAFE-GRPO checkpoints are on ModelScope and Hugging Face; the upgraded SocNav benchmark lives in ABot-Navigation.

---

🌐 Project Page

For an overview of the project, figures, and teaser video, please visit the project page:

👉 Project Page: https://amap-eai.github.io/SocialNav/

---

🔍 Overview

SocialNav is designed to address socially-aware navigation in real-world environments by:

• Combining a VLM-based Brain for high-level semantic and social reasoning
• With a flow-based Action Expert for low-level trajectory generation
• Training on the large-scale SocNav Dataset (7M samples) and evaluating on the SocNav Benchmark

Key components include:

• SocNav Dataset

  • Expert Trajectories Pyramid (ETP)
  • Cognitive Activation Dataset (CAD)

• SocNav Benchmark

  • High-fidelity evaluation built on Isaac Sim + 3DGS
  • 9 large-scale social scenes (parks, streets, offices, campus)
  • Metrics for both navigation performance and social compliance

---

🤗 Models

SAFE-GRPO checkpoints:

Backbone	ModelScope	Hugging Face
Qwen2-VL	SocialNav-Qwen2-VL-SAFE-GRPO	SocialNav-Qwen2-VL-SAFE-GRPO
Qwen2.5-VL	SocialNav-Qwen2.5-VL-SAFE-GRPO	SocialNav-Qwen2.5-VL-SAFE-GRPO

These two checkpoints are the supported path for reproducing the reported results. The Qwen3-VL model code is kept in the repository for reference only.

---

📂 Dataset & Benchmark

• SocNav training dataset: due to data privacy considerations, we have no plans to release it. We appreciate your understanding.
• SocNav benchmark: comprehensively upgraded and released separately in ABot-Navigation.

---

📦 Installation

Requirements

• OS: Linux (recommended) or macOS; GPU inference requires NVIDIA CUDA
• Python: 3.10 / 3.11 (match requirements.txt)
• CUDA: Version compatible with your PyTorch wheels (e.g. cu12)

1. Clone

git clone https://github.com/AMAP-EAI/SocialNav.git
cd SocialNav

2. Virtual environment (optional)

python -m venv .venv
source .venv/bin/activate 

3. PyTorch

Install from pytorch.org for your CUDA version, e.g.:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124

4. Dependencies

pip install -r requirements.txt

This already covers the flow-matching stack (torchcfm, diffusers) and, via its first line, the editable install of the local transformers/ tree described below.

5. Local Transformers (required)

This repo ships a patched Flow Matching tree under transformers/. The SocialNav action expert is wired into Qwen2-VL and Qwen2.5-VL (the two released backbones); a Qwen3-VL implementation is also present for reference. If you did not install requirements.txt, install it explicitly:

pip install -e "./transformers[dev]"   # or: pip install -e ./transformers

6. PYTHONPATH

modeling_qwen3_vl.py imports src.train.sde_with_logprob at module level, and src/train/monkey_patch_forward.py imports the Qwen3-VL module, so the repo root has to be importable even when only Qwen2-VL / Qwen2.5-VL are used. Run from the repo root or set:

export PYTHONPATH="/path/to/SocialNav:${PYTHONPATH}"

For the same reason diffusers is a hard requirement, not an optional one. The Qwen2-VL / Qwen2.5-VL files themselves import the SAFE-GRPO schedules lazily and only need them when grpo_mode=True.

7. Flash Attention (optional)

Install flash-attn-2 per the Qwen2-VL / Qwen2.5-VL docs if you need lower memory; otherwise PyTorch SDPA is fine.

---

📊 Evaluation

Script: utils/citywalker.py — works with either released backbone (Qwen2-VL / Qwen2.5-VL). The architecture is detected from the checkpoint's config.json and the <input_pos*> token ids are resolved from the checkpoint's tokenizer, so no code edits are needed to switch backbones.

Primary metric: mean_angle in metrics_citywalker_<model_type>.csv. Per sample, the script takes the maximum over five steps of the angle (degrees) between predicted and GT waypoint vectors; mean_angle in the CSV is the mean of that value over included samples (by row: categories, overall, and mean). Implementation: compute_sample_metrics and the mean_angle lists in main.

Input (jsonl, one record per line):

Field	Description
images	List of local image paths
messages[0].content	User text, containing the <input_pos1>..<input_pos5> and <input_target> placeholders
messages[1].gt_waypoints	(5, 2)
messages[1].input_waypoints	(6, 2) — 5 history positions plus the goal
messages[1].step_scale	float
messages[1].arrive	[0] or [1]
messages[1].categories	Aligned with TEST_CATEGORIES in the script

Run:

cd /path/to/SocialNav
export PYTHONPATH="$(pwd):${PYTHONPATH}"
CUDA_VISIBLE_DEVICES=0 python utils/citywalker.py \
    --model-path /path/to/SocialNav-Qwen2-VL-SAFE-GRPO \
    --data-path  /path/to/citywalker_test.jsonl \
    --device cuda:0 \
    --flow-steps 5

Outputs: pred_citywalker_<model_type>.jsonl and metrics_citywalker_<model_type>.csv, written to --output-dir (default <model-path>/infer_result_citywalker/).

---

📝 Citation

If you find this project useful in your research, please consider citing (to appear at CVPR 2026):

@article{chen2025socialnav,
      title={SocialNav: Training Human-Inspired Foundation Model for Socially-Aware Embodied Navigation},
      author={Ziyi Chen and Yingnan Guo and Zedong Chu and Minghua Luo and Yanfen Shen and Mingchao Sun and Junjun Hu and Shichao Xie and Kuan Yang and Pei Shi and Zhining Gu and Lu Liu and Honglin Han and Xiaolong Wu and Mu Xu and Yu Zhang},
      journal={arXiv preprint arXiv:2511.21135},
      year={2025}
}