DeReCo: Decoupling Representation and Coordination Learning for Object-Adaptive Decentralized Multi-Robot Cooperative Transport

Kazuki Shibata1*, Ryosuke Sota1*, Shandil Dhiresh Bosch1,2* (*Equal contribution)

Yuki Kadokawa1, Tsurumine Yoshihisa1 , Takamitsu Matsubara1

1 Nara Institute of Science and Technology (NAIST), Robot Learning Laboratory

2 Delft University of Technology (TU Delft)

Youtube

arXiv

Multi-robot cooperative transport of diverse objects

In this study, we address multi-robot cooperative transport of diverse objects, where multiple robots cooperatively transport an object to a goal position across diverse shapes and physical properties. Each robot has local observations exclude other robots’ observations as well as privileged information about object-dependent properties such as shape, mass, and friction.

Challenges

Representation learning under partial observability
- Robots must infer object-dependent representations from limited sensory inputs.
Coordination learning under MARL non-stationarity
- Each agent’s policy evolves during training, making the environment non-stationary from the perspective of the other agents.

Joint optimization of representation and coordination learning

A typical approach jointly optimizes object-dependent representation and coordinated policies in an end-to-end manner while randomizing object shapes and physical properties.

Limitation of joint optimization

The joint optimization tightly couples representation and coordination learning, introducing bidirectional interference.
Inaccurate representations under partial observability destabilize coordination, while MARL non-stationarity degrades representation learning, resulting in sample-inefficient and unstable training.

Idea

The core idea is to decouple object-dependent representation learning under partial observability from coordination learning in MARL. We first learn stable coordination policies by providing privileged object information during training, which yields reliable object-dependent representations. We then learn object-dependent representations from local observations on top of the stabilized coordination policies.

Decoupling Representation and Coordination Learning (DeReCo)

This study proposes DeReCo, a novel MARL framework that decouples representation and coordination learning for object-adaptive multi-robot cooperative transport, improving sample efficiency and robust generalization across objects and transport scenarios. DeReCo adopts a three-stage training strategy: (1) centralized coordination learning with privileged object information, (2) reconstruction of object-dependent representations from local observations, and (3) progressive removal of privileged information for decentralized execution. This decoupling mitigates representation-coordination coupling and enables stable and sample-efficient learning.