We are working with machine learning based methodologies to develop data-driven system identification and control techniques with applications to aerospace systems such as UAVs and spacecrafts. Our approach lets us linearize highly non-linear dynamical systems and thus provide a robust linearization scheme compared to traditional techniques such as Taylor series expansion.

We are working on developing a 3D reconstruction pipeline for uncooperative satellites using deep learning-based 3D reconstruction methods (Neural Radiance Fie…We are working on developing a 3D reconstruction pipeline for uncooperative satellites using deep learning-based 3D reconstruction methods (Neural Radiance Fields). Our team creates synthetic datasets using Isaac Sim alongside real-world mock satellite setups to train and validate these models across a variety of orbital trajectories and lighting conditions. We are tackling key challenges such as optimal viewpoint sampling, dynamic reconstruction of tumbling satellites, and improving reconstruction fidelity in realistic scenarios through background removal and photometric correction techniques.

We are exploring on how multi-agent reinforcement learning (MARL) can be used to enhance the autonomous inspection of spacecraft. Building on prior work in autonomous spacecraft inspection, we focused on the design and validation of reward functions for MARL inspection when the objective is not just reaching predefined inspection points, but collecting useful imagery for 3D reconstructions. The approach taken in this work allows spacecraft to decide when and where images should be captured. By then feeding trajectories and camera poses into NVIDIA Isaac Sim, we were able to recreate the scene and trajectories from the RL environment and subsequently capture photorealistic inspection images. These images and the associated camera poses were fed into the 3D reconstruction algorithms COLMAP and Instant-NGP, and the resulting reconstructions were then used to validate and improve the reward function design. As a result of this work, we formulated a flexible framework and reward functions that led to the successful reconstruction of the targets used for validation.

We have designed several adaptive and optimal control techniques for close proximity operations near asteroids. We are currently working designing navigation and path planning techniques for spacecraft swarms near asteroids.