Artificial Intelligence

   

Artificial intelligence is increasingly becoming part of the connective tissue of our autonomy stack. We apply machine learning and deep learning at multiple levels of the perception–decision–action loop across all three mission domains. In planetary robotics, deep neural networks trained on photorealistic simulation data (closing the sim-to-real gap efficiently) detect and localize sample tubes and science targets, while multimodal transformer-based architectures fuse RGB, depth, and thermal imagery for semantic terrain segmentation. In orbital robotics, AI-driven anomaly detection and 4D shape reconstruction underpin our autonomous inspection pipeline. In robotic telescopes, our long-standing work with the BOOTES network and the GLORIA distributed-telescope platform has produced scheduling systems that autonomously plan and dispatch observation sequences in response to real-time alerts—a key capability for multi-messenger astronomy follow-up. We also maintain an active interest in representation learning, sim-to-real transfer, and science autonomy— equipping robots to evaluate the scientific merit of discovered targets and prioritize independently.

KEY CAPABILITIES
Sim-to-real transfer learning GPU-accelerated networks trained on photorealistic renders achieve real-world sample detection without labelled field data	Multimodal terrain segmentation and 3D reconstruction OmniUnet fuses RGB, depth, and thermal modalities via a Swin-based architecture for safe rover traversal.
Autonomous telescope scheduling GlSch and GLORIA-era planners dynamically schedule observations across a distributed network in response to transient alerts.	Science autonomy AI classifiers rank science targets by interest metrics, enabling rover on-board prioritization without ground-uplink latency.

 

IN THE LAB & IN THE FIELD

 

SELECTED PUBLICATIONS
01	High-fidelity 3D reconstruction for planetary exploration Martínez-Petersen, A., Gerdes, L., Rodríguez-Martínez, D. and Pérez-del-Pulgar, C.J. IEEE Conference on Artificial Intelligence (CAI) Special Session on "AI for Space Exploration," Granada, Spain, (2026).
02	OmniUnet: A Multimodal Network for Unstructured Terrain Segmentation on Planetary Rovers Using RGB, Depth, and Thermal Imagery Castilla-Arquillo R., Pérez-del-Pulgar C.J., Gerdes L., García-Cerezo A., Olivares-Mendez M. International Conference on Space Robotics (iSpaRo), Sendai, Japan (2025)
03	Hardware-Accelerated Mars Sample Localization Via Deep Transfer Learning From Photorealistic Simulations Castilla-Arquillo, R., Pérez-del-Pulgar, C. J., Paz-Delgado, G. J., and Gerdes, L.  IEEE Robotics & Automation Letters, 7 (4), 12555-12561 (2022)
04	Samples detection and retrieval for a sample fetch rover, Mantoani, L., Castilla-Arquillo, R., Paz Delgado, G. J., Perez-del Pulgar-Mancebo, C. J., & Azkarate, M. 16th Symposium on Advanced Space Technologies in Robotics and Automation (2022).
05	Very-High-Frequency Oscillations in the Main Peak of a Magnetar Giant Flare Castro-Tirado A.J., Østgaard N., Göğüş E., Sánchez-Gil C., et al.  Nature, 600(7890), 621–624 — 2021
06	GlSch: Planificación de Observaciones en la Red de Telescopios GLORIA López-Casado M.C., Pérez-del-Pulgar C.J., Muñoz V.F., Castro-Tirado A. Revista Iberoamericana de Automática e Informática Industrial (RIAI), 15(3), 339–350 — 2018

OPEN-SOURCE SOFTWARE & DATASETS
	OmniUnet Code for training and executing OmniUNet, a multimodal neural network based on transformers designed for semantic segmentation of images that combine color, depth, and thermal data github.com/spaceuma/OmniUnet  >
	Mars Sample Localization Code associated with the Hardware-accelerated Mars Sample Localization via deep transfer learning from photorealistic simulations paper github.com/spaceuma/MarsSampleLocalization >

SEE ALSO