Adaptive Multi-Robot Social Navigation via Joint Human-Response Prediction and Interaction-Aware Replanning

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

TRACER predicts how people respond to joint robot actions, executes a short plan, and updates persistent beliefs from the observed response before replanning.

• Predict joint-action responses: individual-robot effects and non-additive pairwise interactions.
• Remember interactions: identity-bound beliefs updated from executed-prefix observations.
• Replan online: response-aware candidate evaluation and synchronized execution.

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 tracer.demo --output outputs/demo

The CPU example trains a small response model on generated data, then runs seven decision windows with persistent belief updates. It writes an animation, a checkpoint and the decision trace.

See the demo guide for inputs, outputs and interpretation.

Method

Experiments and results

Experiment	Reported result
Controlled joint-action response	Pair-effect RMSE decreases 23.94% relative to the capacity-matched additive predictor
Identity and memory	Persistent beliefs evaluated under ID/OOD identity and memory ablations
SocialGym2 navigation	Collision-free completion rises from 22.5% to 31.6% against GoAlone over 320 paired episodes

The result index identifies the protocols and status of 14 retained CSV files. Commands and required inputs are in the experiment guide.

Models

A lightweight synthetic demonstration checkpoint is available. The model guide describes its scope and the additional assets required for formal experiments.

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

tracer/        Response models, belief updates and planning
tracer/legacy/ Archived-data compatibility modules
configs/       Indexed formal, ablation and development protocols
tests/         CPU regressions and asset-dependent checks
results/       Selected tables and paired statistics
provenance/    File origins and checksums

Citation

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

@misc{tracer_manuscript,
  title = {TRACER: Adaptive Multi-Robot Social Navigation via Joint Human-Response Prediction and Interaction-Aware Replanning},
  author = {Hu, Lan and Liwang, Minghui and Zhu, Wenbo and Yi, Xinlei and Gong, Wei and Hong, Yiguang and Hosseinalipour, Seyyedali},
  year = {2026},
  eprint = {2609.18776},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2609.18776}
}