SatBench: Satellite Coordination Benchmark Based on Multi-Agent Reinforcement Learning

ICLR 2026 Conference Submission14804 Authors

Keywords: Satellite tasking, Multi-agent reinforcement learning

TL;DR: SatBench, a realistic benchmark for evaluating MARL in coordinated satellite tasking for Earth observation missions.

Abstract: We introduce SatBench, a realistic and flexible benchmark for coordinated satellite tasking in Earth observation (EO) missions. In this setting, multiple satellites potentially in different orbits must coordinate to capture images of ground targets under time constraints. While real-world satellites involve complex sensing and control dynamics, SatBench abstracts these into an agent-based learning framework with a standardized, user-friendly interface. This enables the application of reinforcement learning (RL) methods while preserving critical realism in the environment. SatBench supports a diverse set of configurable scenarios that capture varying EO requirements, including different satellite formations, target distributions, and prioritization schemes. We evaluate representative multi-agent reinforcement learning (MARL) algorithms across these scenarios, highlighting key challenges such as coordination under temporal coupling and scalability. SatBench aims to foster progress in both autonomous satellite coordination and the development of more robust, generalizable MARL methods. SatBench is available at https://anonymous.4open.science/r/SatBench-43DB.

Primary Area: datasets and benchmarks

Submission Number: 14804