SeqWM: Sequential World Model for Multi-Robot Cooperation (ICLR'26)

The official implementation of the paper Empowering Multi-Robot Cooperation via Sequential World Models, which published on ICLR 2026.

Overview

To address the difficulty of applying model-based reinforcement learning (MBRL) to multi-robot systems, we propose the Sequential World Model (SeqWM). This framework decomposes complex joint dynamics by using independent, sequentially-structured models for each agent. Planning and decision-making occur via sequential communication, where each agent bases its actions on the predictions of its predecessors. This design enables explicit intention sharing, boosts cooperative performance, and reduces communication complexity. Results show SeqWM outperforms state-of-the-art methods in simulations and real-world deployments, achieving advanced behaviors like predictive adaptation and role division.

⚙️ Installation

bash
conda create -n seqwm python=3.8 -y
conda activate seqwm
pip install -r requirements.txt

If you encounter issues when installing Isaac Gym, Bi-DexHands or MQE please refer to Isaac Gym, Bi-DexHands and MQE.

🚀 Quick Start

For training, please run:

bash
python examples/train.py --load_config configs/dexhands/ShadowHandBottleCap/seqwm/config.json

You can modify the configuration in configs/{env_name}/{task_name}/seqwm/config.json to customize the training process.

📈 Curves

SeqWM consistently outperforms state-of-the-art baselines in both Bi-DexHands and Multi-Quad environments.

🎥 Demos

👐 Bi-DexHands

These GIFs showcase SeqWM’s ability to solve complex bimanual manipulation tasks.

🤖 Multi-Quadruped

In the Multi-Quad environment, SeqWM supports scalable cooperation among 2–5 quadruped robots.

🌍 Sim2Real Deployment

SeqWM has also been successfully deployed on real Unitree Go2-W robots, confirming effective sim-to-real transfer.

🧩 Advanced Cooperative Behaviors

As shown below, the robots exhibit predictive yielding and temporal alignment: some agents slow down in front of the gate (observable as troughs in their x-axis velocity commands), while others accelerate and pass through first (peaks in velocity commands).
This wave-like pattern across agents reflects turn-taking and priority management, enabling smooth passage without collisions even in highly constrained environments.

🙏 Acknowledgement & 📜 Citation

Our code is built upon HARL, TDMPC2 and M3W. We thank all these authors for their nicely open sourced code and their great contributions to the community.

If you find our research helpful and would like to reference it in your work, please consider the following citations:

bibtex
@inproceedings{
    zhao2026seqwm,
    title     = {Empowering Multi-Robot Cooperation via Sequential World Models},
    author    = {Zijie Zhao and Honglei Guo and Shengqian Chen and Kaixuan Xu and Bo Jiang and Yuanheng Zhu and Dongbin Zhao},,
    booktitle = {The Fourteenth International Conference on Learning Representations},
    year      = {2026},
    url       = {https://openreview.net/forum?id=IvUM6UwYCJ}
}