A Socially-Aware Generative Engine for Heterogeneous Multi-Agent Navigation

Project page · Quick start · Experiments · Models · Results · Paper · 中文

SAGE jointly generates robot plans and surrounding-agent trajectories, then refines robot motion with safety and social guidance.

• Encode heterogeneous interactions: directed, role-aware graph representations.
• Generate joint futures: conditional diffusion for robot plans and entity trajectories.
• Refine robot motion: differentiable safety, social, kinematic and goal guidance at inference.

Quick start

Use Python 3.12 and follow the installation guide to create an environment. From this repository:

python -m pip install -e ".[dev]"
python -m sage.demo --output outputs/demo

The CPU example refines a crossing-scene robot trajectory while keeping entity predictions fixed. It writes an animation, a final plot and an energy trace.

See the demo guide for inputs, outputs and interpretation.

Method

Experiments and results

Setting	Retained results
ETH/UCY	Prediction, guidance and ablation records using a target-as-robot proxy
Stanford Drone Dataset	Navigation results using semantic annotations
Synthetic navigation	Guidance trade-offs and robot-count scaling

The result index identifies the protocols and status of 44 retained CSV files. Commands and required inputs are in the experiment guide.

Qualitative example

Three robots (R1–R3) navigating a Stanford Drone Dataset crossing (hyang_0). Entity predictions are shown with the robot plans before and after guidance; see the paper for the legend. Scene imagery belongs to the Stanford Drone Dataset and follows its terms.

Models

Six original goal-conditioned research checkpoints are available for ETH/UCY and SDD.

python tools/download_models.py

Downloads are verified against the release checksum. See model details for checkpoint identity and use.

Development

python tools/check_repository.py
python -m pytest -q

CPU checks run on Linux and Windows. Tests requiring archived experiment outputs or CUDA report their prerequisites through explicit skips. See verification.

Repository layout

sage/          Models, guidance, data adapters and experiment modules
configs/       Dataset and ablation configurations
results/       Selected numerical records
docs/          Setup, examples and experiment guides
provenance/    File origins and checksums

The top-level models/, dataset/, environment/, evaluation/, utils/, mid.py, main.py, process_data.py and visualisation.py are MID/Trajectron++ compatibility code kept in place because archived checkpoints and processed data reference these module paths. See third-party notices.

Citation

Please cite the paper when using SAGE in research. BibTeX · CITATION.cff.

@misc{sage_manuscript,
  title = {SAGE: A Socially-Aware Generative Engine for Heterogeneous Multi-Agent Navigation},
  author = {Hu, Lan and Liwang, Minghui and Zhu, Wenbo and Yi, Xinlei and Hong, Yiguang and Wang, Xianbin and Jiao, Zhenzhen and Hosseinalipour, Seyyedali},
  year = {2026},
  eprint = {2607.16619},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2607.16619}
}