Communication-Aware Heterogeneous Graph Learning for Decentralized Multi-Human Multi-Robot Task Allocation

Anonymous Author(s)

[Benchmark(Anonymous GitHub)]

ICRA 2027

Abstract

Multi-human multi-robot (MH-MR) teams combine robotic autonomy with human expertise, but effective task allocation requires coordinating scarce, dynamically available human support with distributed robot execution. Limited robot-robot and human-robot communication further complicates this coupling by delaying information exchange and supervisory intervention. We introduce CommHG, a communication-aware heterogeneous graph learning framework for decentralized MH-MR task allocation. CommHG represents coupled human-robot-task interactions through local graphs conditioned on information availability and age. Learned communication actions enable robots to decide when and which operator to query, and when to share information with peers. Allocation and communication are jointly optimized through cooperative multi-agent reinforcement learning, allowing the team to acquire useful information while managing limited communication and supervisory resources. We also introduce a benchmark integrating heterogeneous humans and robots, dynamic tasks and operational states, and constrained robot-robot and bidirectional human-robot links. Experiments across heterogeneous teams with up to 16 robots, 6 humans, and 112 tasks show that CommHG improves timely weighted mission completion as coordination scale increases, with an approximately 10-point advantage over the strongest competing baseline in the Large scenario.

Challenges in MH-MR Coordination Under Communication Constraints

Fig. 1a Human–robot–task interaction flow

Fig. 1b MH-MR collaboration under limited communication Human–robot–task interaction flow

A coupled human-robot-task interaction flow: a task determines the need for assistance, an operator's support changes its execution time and reliability, and allocating that operator reduces the support available to other robots. Human proficiency, evolving fatigue, and ongoing supervisory commitments further change the value and availability of assistance. Treating a human simply as another independent task executor misses these dependencies.

Framework Overview

Fig. 2 Overview of CommHG for decentralized MH-MR task allocation under limited communication.

(a) Each decision maker uses its own state, public task attributes, and received records, together with local communication resources.

(b) Typed encoders and relation-specific aggregation represent human supervision, robot execution, and task demand. Local graphs retain information availability and age: crosses indicate unavailable records, and clocks indicate delayed information.

(c) Role-specific policies jointly learn task allocation, peer broadcasting, operator queries, and human supervisory allocation. Robots decide when and which operator to query, while humans decide which robot to assist. A centralized critic uses global state during PPO training, while execution relies on local observations. Actions affect the graph at subsequent decisions through locally observed changes and delivered messages.

Benchmark and Simulation Environment

Fig. 3 Overview of the MH-MR benchmark environment: human operators, UAVs, UGVs, terrain obstacles, and three task types — general, UAV-only, and UGV-only.

Rendered in Blender

⬇️ Rendered simulations for every method on the Large map — 6 human operators, 8 UAVs, 8 UGVs, and 112 tasks ⬇️

Greedy

Mean Score: 0.69

CommHG

Mean Score: 0.89

MAPPO

Mean Score: 0.70

IPPO

Mean Score: 0.64

InforMARL

Mean Score: 0.79

ATA-HRL

Mean Score: 0.72

Experiments

Fig. 4 Mean TWCR (a) Baseline comparison across Small, Medium, and Large scenarios. (b) Ablations in the Medium scenario under increasing communication limitation. Numbers above bars indicate mean values.

Fig. 5 Training curves over ten seeds. They are trained on medium difficulty, evaluated on three difficulties

Google Sites

Report abuse