New MARL Reward Function Improves Autonomous Satellite Inspection

A team led by Patrick Quinn at the AIAA SCITECH 2026 Forum presents a simulation-based method for validating reward functions in multi-agent reinforcement learning (MARL) for on-orbit inspection. The core innovation is a generalized reward function that moves beyond the traditional approach of using a finite set of predetermined inspection points. Instead, the function evaluates any number of images captured at arbitrary locations, using 3D reconstructions of the inspected object as a reference. This allows the trained MARL agents to have full autonomy over when to capture images during the inspection mission, adapting their strategy based on real-time data rather than following a fixed plan.

The paper (arXiv:2607.01367) includes 13 pages and 6 figures, and integrates a published correction. The authors derive key insights into best practices not only for the specific MARL inspection task but also for broader orbital inspection scenarios outside the MARL context. By validating the reward function through simulation, the work demonstrates a more flexible and potentially more efficient method for controlling groups of inspection satellites, which could reduce the need for ground-based command and increase the robustness of autonomous space operations.