Research map / Architectures and efficiency / RNNs and normalization

Towards Realistic Earth-Observation Constellation Scheduling: Benchmark and Methodology

NeurIPS 2025 · Architectures and efficiency › RNNs and normalization · approach: Distributed LLM inference systems · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces AEOS-Bench, a large-scale physics-simulated benchmark (16,410 scenarios, 3,907 satellite assets) for agile Earth observation constellation scheduling, and AEOS-Former, a Transformer scheduler with a constraint module that masks cross-attention, trained with simulation-based iterative learning.

Problem

Agile Earth observation satellite (AEOS) constellation scheduling is hard to solve at large scale, under dynamic environments and strict physical constraints (energy, FOV, attitude control, time windows), and existing methods simplify these and lack a common realistic benchmark.

Approach

A Transformer encoder-decoder scheduler whose cross-attention mask is derived from an internal constraint module that predicts per satellite-task feasibility logits and attitude adjustment time, combined with simulator-in-the-loop iterative training that adds high-scoring simulated schedules back to the training set.

Key results

• On the test split, AEOS-Former achieves CS 6.28 versus 7.33 for MSCPO-SHCS (about 16.7% better) and lower power consumption (40.91 Wh vs 149.20 Wh), though MSCPO-SHCS has higher completion rate (19.44% vs 19.25%).
• On val-unseen, AEOS-Former reaches 35.42% completion rate and 68.99 Wh power consumption, versus 35.35% and 140.83 Wh for MSCPO-SHCS.
• Ablation on the test split: removing both the constraint module and iterative training gives CS 9.31; the full model gives CS 6.28.

Measured effects reported in the paper

Surprising findings

• On the test split, MSCPO-SHCS has a slightly higher completion rate (19.44%) than AEOS-Former (19.25%), and on val-seen the iterative training alone gives higher completion rate than the full model, showing a trade-off between completion rate and power consumption.

Limitations

• Each task is represented as a single location point; area-based imaging tasks are not supported.
• Constraint module is trained on approximate labels because individual satellite contribution to multi-satellite tasks is not directly labeled.
• Completion rate is not uniformly best; the model trades completion rate against power consumption.

Experimental setup

Datasets: AEOS-Bench train split (16,218 trajectories) simulator-generated scenarios N2YO and Gunter's Space Page real satellite data (test split)

Benchmarks: AEOS-Bench val-seen AEOS-Bench val-unseen AEOS-Bench test

Metrics: comprehensive score completion rate partial completion rate weighted completion rate turn-around time power consumption

Baselines: Random HAAL REDA MSCPO-SHCS

Builds on: Transformer Basilisk

Closest papers by meaning

• COSMIC: Compress Satellite Image Efficiently via Diffusion Compensation (NeurIPS 2024 · similarity 0.80)
• Resource-Aware Parameter-Efficient Model Adaptation for Onboard High-Dimensional Data (NeurIPS 2026 · similarity 0.80)
• REOBench: Benchmarking Robustness of Earth Observation Foundation Models (NeurIPS 2025 · similarity 0.80)
• DAWP: A framework for global observation forecasting via Data Assimilation and Weather Prediction in satellite observation space (NeurIPS 2025 · similarity 0.79)
• Mars-Bench: A Benchmark for Evaluating Foundation Models for Mars Science Tasks (NeurIPS 2025 · similarity 0.79)
• Earth-Agent: Unlocking the Full Landscape of Earth Observation with Agents (ICLR 2026 · similarity 0.79)