SeqWM

ICLR 2026

SeqWM: Empowering Multi-Robot Cooperation via Sequential World Models

Zijie Zhao1,2, Honglei Guo2, Shengqian Chen2, Kaixuan Xu1,2, Bo Jiang2,1, Yuanheng Zhu2,1,*, Dongbin Zhao2,1

1 School of Artificial Intelligence, University of Chinese Academy of Sciences

2 SKL-MAIS, Institute of Automation, Chinese Academy of Sciences

* Corresponding author.

Code arXiv PDF OpenReview

Plan with your teammates’ predictions.
Sequential world models turn shared intentions into coordinated actions.

Abstract

Model-based reinforcement learning (MBRL) has achieved remarkable success in robotics due to its high sample efficiency and planning capability. However, extending MBRL to physical multi-robot cooperation remains challenging due to the complexity of joint dynamics. To address this challenge, we propose the Sequential World Model (SeqWM), a novel framework that integrates the sequential paradigm into multi-robot MBRL. SeqWM employs independent, autoregressive agent-wise world models to represent joint dynamics, where each agent generates its future trajectory and plans its actions based on the predictions of its predecessors. This design lowers modeling complexity and enables the emergence of advanced cooperative behaviors through explicit intention sharing. Experiments on Bi-DexHands and Multi-Quadruped demonstrate that SeqWM outperforms existing state-of-the-art model-based and model-free baselines in both overall performance and sample efficiency, while exhibiting advanced cooperative behaviors such as predictive adaptation, temporal alignment, and role division. Furthermore, SeqWM has been successfully deployed on physical quadruped robots, validating its effectiveness in real-world multi-robot systems.

Method

Sequential prediction. Cooperative planning.

Decompose complex joint dynamics into independent agent-wise world models, connected through predictions of future trajectories.

SeqWM framework for sequential world modeling and cooperative action planning
SeqWM framework. Each agent conditions its world model and planning on the predicted trajectories and actions of its predecessors. Click any figure to view it at full resolution.

  1. 01 · Predict

    Each agent models its own dynamics, conditioned on the predictions of preceding agents.

  2. 02 · Share

    Predicted trajectories communicate intentions explicitly along the agent sequence.

  3. 03 · Plan

    Each agent performs multi-step lookahead in its own action space while accounting for its teammates.

Experiments

Cooperation across tasks.

SeqWM is evaluated on Bi-DexHands and Multi-Quadruped, with deployment on physical Unitree Go2-W robots.

Learning curves comparing SeqWM against model-based and model-free baselines on Bi-DexHands and Multi-Quad
Benchmark performance. SeqWM improves overall performance and sample efficiency across the evaluated multi-robot tasks.

Real-world cooperation

Sim-to-real transfer to a team of two Unitree Go2-W robots.

Gate · Coordinated passage
PushBox · Cooperative transport
Shepherd · Cooperative herding

Bi-DexHands

Coordinated object transfer and precise bimanual manipulation.

Over
CatchAbreast
CatchOver2Underarm
BottleCap
Pen
Scissors

Multi-Quadruped

Cooperation among two to five quadruped robots in simulation.

Gate · 2 robots
Gate · 3 robots
Gate · 4 robots
Gate · 5 robots
PushBox
Shepherd

Advanced cooperative behaviors

Learning when to lead and when to yield.

Shared predictions support predictive adaptation, temporal alignment, and role division.

Five-robot gate-passage trajectories and velocity commands showing sequential passage and predictive yielding
Coordinated gate passage. Some agents slow down before the gate while others accelerate and pass first. The wave-like velocity patterns reflect turn-taking and temporal alignment in a constrained environment.

Citation

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}
}