MIMIC-D: Multi-modal Imitation for MultI-agent Coordination with Decentralized Diffusion Policies

In this project we introduce MIMIC-D, a CTDE (Centralized Training, Decentralized Execution) framework that learns decentralized diffusion policies from multi-agent expert demonstrations to recover diverse, coordinated behaviors without explicit inter-agent communication.

Paper
Website

Overview

Many real-world multi-agent tasks have multiple valid coordination modes (e.g., pass-left vs pass-right) and cannot assume reliable centralized planners or explicit communication. MIMIC-D trains policies jointly with full information, then executes with only local observations, enabling implicit coordination while preserving multi-modality in the learned behaviors. We validate MIMIC-D in multiple simulation environments and on a bimanual hardware setup with heterogeneous arms (Kinova3 + xArm7).

Project Organization

• dependencies/ — conda environment file (it may be easier to simply install dependencies as you go)
• lift/ — simulated two-arm pot lifting experiment in robosuite
• lift_hardware/ — two-arm pot lifting experiment on Kinova3 and xArm7 on hardware
• three_agent_road/ — three agent road crossing environment
• two_agent_swap/ — two agent swap environment
• docs/ — all the elements to build the project website

Getting Started (fill in after release)

1. TODO: environment setup (conda, CUDA/cuDNN, PyTorch version, robosuite, etc.)
2. TODO: data preparation (where to download / how to format expert demos)
3. TODO: training (commands & key flags)
4. TODO: sampling / evaluation (receding-horizon execution, metrics, plotting)

Citation

If you use MIMIC-D, please cite:

@article{dong2025mimic,
  title={MIMIC-D: Multi-modal Imitation for MultI-agent Coordination with Decentralized Diffusion Policies},
  author={Dong, Dayi and Bhatt, Maulik and Choi, Seoyeon and Mehr, Negar},
  journal={arXiv preprint arXiv:2509.14159},
  year={2025}
}

Acknowledgments

Our diffusion transformer architecture is largely based on the AlignDiff code.