Are We Ready to Enable Satellite Autonomy Through on Board Image Registration?

Enabling autonomous CubeSats requires shifting image-based decision-making from ground stations directly to on-board Commercial-Off-The-Shelf (COTS) hardware to overcome the unsustainable downlink bottlenecks typical of Low Earth Orbit (LEO) missions. A critical yet largely neglected aspect of this transition is Image Registration (IR), which is central for multi-temporal analysis, change detection, and multimodal data fusion. Despite its pivotal role, existing IR methods are designed and evaluated on unconstrained computing platforms, making their readiness for on-board autonomous operations unclear due to the lack of the assessment of latency-accuracy-energy trade-offs. In this paper, we show that state-of-the-art ''out-of-the-box'' IR algorithms fail to reconcile the latency, accuracy, and energy constraints inherent to COTS-based satellite platforms. We present STAR-Bench , an open-source framework for systematic and reproducible characterization of IR methods on on-board hardware across diverse sensing modalities, terrains, and deformation regimes. Our characterization reveals a fundamental specialization gap: lightweight methods meet latency constraints but fail under multimodal conditions; while, robust methods incur up to an order-of-magnitude higher energy costs, precluding effective autonomy. To capture system-level trade-offs, we introduce Quality-Adjusted Cost (QAC), a metric to quantify the energy required to achieve a target accuracy level. Guided by this analysis, we propose ETNA , a mission-aware framework to dynamically orchestrate heterogeneous IR pipelines. To address the challenging multimodal scenario, we perform a hardware-software co-design of an FPGA-accelerated solution. ETNA reduces latency by up to 1.7× in Synthetic Aperture Radar (SAR)-Optical scenarios and improves QAC by up to 12.03× over all Pareto-optimal baselines, while also improving registration accuracy. Our results demonstrate that satellite autonomy cannot be achieved by deploying existing IR algorithms unchanged; instead, IR must be explicitly co-designed with COTS hardware and mission constraints.