Workshop Program

AI models launched on spacecraft are often frozen at deployment, yet sensing conditions and operational demands in orbit change. In-orbit updates are hard because the standard loop (e.g., downlinking new in-orbit data, such as full-resolution images, labelling them on Earth, retraining models on Earth, uplinking new weights) is constrained by bandwidth, power, thermal limits, and short operations windows. We show that meaningful in-mission adaptation is possible under these constraints and introduce Telemetry-First In-orbit Fine-Tuning (TFiT), an operations-first framework for fine-tuning AI models in space. Telemetry denotes compact kilobyte-scale non-image records downlinked from the spacecraft, including logits, timestamps, geolocation, and spacecraft status. TFiT avoids routine image downlink: labels are produced on Earth by fusing telemetry with Earth-based context sources, while thumbnails are downlinked for human review only when label evidence is low-confidence or conflicting. Crucially, in-orbit data remains onboard and only approved labels are uplinked for bounded on-orbit updates. We demonstrate TFiT on the flight-operated 6U SpIRIT nanosatellite by executing an in-orbit update of an onboard cloud-detection model (OrbitBaseline to OrbitAdapted) and evaluating pre/post-update behavior with a same-image audit protocol. Under the recovered same-image audit, OrbitAdapted improves ROC AUC by +0.510 and F1 score by +0.871 over the pre-update baseline. These results demonstrate in-mission adaptation without routine full image downlink, widening the practical scope of AI in space.

Image-based surface reconstruction and characterization are crucial for missions to small celestial bodies (e.g., asteroids), as it informs mission planning, navigation, and scientific analysis. Recent advances in Gaussian splatting enable high-fidelity neural scene representations but typically rely on a spherical harmonic intensity parameterization that is strictly appearance-based and does not explicitly model material properties or light-surface interactions. We introduce AstroSplat, a physics-based Gaussian splatting framework that integrates planetary reflectance models to improve the autonomous reconstruction and photometric characterization of small-body surfaces from in-situ imagery. The proposed framework is validated on real imagery taken by NASA's Dawn mission, where we demonstrate superior rendering performance and surface reconstruction accuracy compared to the typical spherical harmonic parameterization.

Identifying Martian minerals from sparse hyperspectral observations is fundamentally constrained by extreme label scarcity and the absence of large-scale hyperspectral pretraining. To this end, we introduce LUMEN (Language-guided Unified Memory for ENcoding hyperspectral images), a memory-augmented cross-modal framework that formulates hyperspectral recognition tasks as structured relational alignment between spectral embeddings and language-derived semantic prototypes. LUMEN employs an ultralight 3D spectral encoder to distill spatial–spectral structure into a compact latent manifold, which is softly aligned to textual embeddings through a learnable bilinear projection. To reconcile linguistic priors with spectral evidence, semantic prototypes are refined via learnable offsets and a confidence-aware gating mechanism that adaptively regulates the contribution of language priors and visual memory. A unified momentum-updated memory bank further stabilizes learning by integrating exponentially smoothed visual representations with language-anchored prototypes, reducing cross-modal drift. Unlike prior language-guided or CLIP-style approaches, LUMEN explicitly enforces eigenspace-level cross-modal consistency through a spectral relational distillation objective that promotes global manifold alignment beyond pointwise embedding matching. Extensive evaluations of three Martian hyperspectral datasets show that LUMEN consistently exceeds strong transformer-based approaches while requiring fewer parameters and reduced training time.

Deploying advanced cognitive capabilities on nanosatellites faces extreme Size, Weight, and Power (SWaP) constraints, cosmic radiation, and software rigidity. Current solutions relying on power-hungry data-center GPUs, heavy cloud-native virtualization, or rigid monolithic cold-spares are physically and operationally unviable for 10W-class nodes. In this paper, we propose the Space Edge AI Computer (EAC), a deeply decoupled 12W/703.5g COTS-centric payload engineered for the nanosatellite SWaP-C sweet spot. Hardware-wise, its physical M.2 decoupling allows for agile pre-launch accelerator swaps to resist chip obsolescence, while a mixed-reliability, Open-Collector (OC)-driven intervention mechanism provides granular intra-module storage switching to prevent entire-board failure. Software-wise, we design AIOS, a zero-virtualization spaceborne middleware that eschews heavy containers for OS-level POSIX process derivation, achieving extreme reconfiguration elasticity over weak telemetry links. We report a flawless initial 124-day bare-LEO flight validation campaign. Our in-orbit evaluations validated the first bare-LEO configurable SLM and RAG architecture for medical consultation (achieving a 7.7× latency reduction), concurrent heterogeneous vision pipelines via a strict isolated-frame protocol, and a pioneering feasibility study of an offline Chain-of-Thought (CoT) reasoning model (DeepSeek-R1-Distill) under strict power envelopes. This establishes a pragmatic, flight-validated foundation for future distributed spaceborne autonomous agents.

Synthetic Aperture Radar (SAR) imagery enables all-weather, day-and-night Earth observation; however, it remains difficult to interpret due to speckle noise and other intrinsic imaging artifacts. Sentinel-1 (S1) constitutes one of the most widely used spaceborne SAR missions, offering systematic global coverage, high temporal resolution, dual-polarization imaging, and free data availability. Among S1 modes, Stripmap (SM) provides the highest resolution, yet speckle noise and spatial constraints often hinder applications requiring finer spatial detail. This motivates the need for effective image enhancement strategies. In this work, we propose a self-supervised enhancement framework for S1 SM imagery based on azimuth subaperture decomposition. The method exploits the physical consistency between subaperture reconstructions and the corresponding full-aperture image to generate paired training data without external sensors, simulated ground truth, or multi-temporal stacks. The proposed framework integrates single- and multi-frame learning and incorporates an iterative inference scheme that progressively refines image quality. Experiments on real S1 SM data show that the proposed approach consistently outperforms the widely adopted self-supervised deep learning baseline MERLIN, in terms of PSNR and SSIM, while MERLIN attains higher ENL, highlighting a trade-off between structural fidelity and speckle smoothing. Overall, the results demonstrate that subaperture-based supervision provides a physically grounded, reproducible, and operationally viable approach for SAR image enhancement using S1 data.