Workshop 1:  TraML Meets AI in Space

Dr. Di Wu Presentation

Abstract:  Benchmarking large language models in space engineering presents unique challenges, particularly regarding model evaluation and interpretability in high-stakes environments. This work introduces an innovative framework based on autoencoder architectures to systematically assess and benchmark AI models used in space applications. By extracting latent representations through optimized autoencoder pipelines, our approach elucidates inherent model behaviors and performance nuances across various architectures. Detailed case studies demonstrate how this method enhances simulation accuracy and risk assessments in complex space scenarios. The framework bridges the gap between deep learning and practical engineering challenges, providing a unified strategy for model evaluation. We discuss both the theoretical foundations and empirical validations of our approach, highlighting its potential to drive advancements in autonomous system design and decision-making for space engineering missions.

Dr. Di Wu:  Di Wu is a faculty in the Aerospace Engineering Department of the Embry-Riddle Aeronautical University and the founder and director of XDLab. He was a postdoc associate in AeroAstro ARCLab at MIT. His work roots in perturbation theory and dynamical system with a broad interest in integrating optimization, ML/AI, and control theory into understanding space situational awareness and space sustainability. He has been actively collaborating with students and professionals all over the world.

Dr. Amit Jain Presentation

Abstract:  Control policies that generalize across dynamically distinct regimes present significant challenges in complex systems like spacecraft mission phases—from orbit-raising to rendezvous. While reinforcement learning (RL) has shown promise in individual control tasks, existing approaches often require separate policies for different dynamical regimes, limiting adaptability. This work introduces a transformer-based RL framework that aims to unify control across varying dynamical contexts through a single policy architecture. Building on proximal policy optimization (PPO), our framework replaces conventional recurrent networks with a Gated Transformer-XL (GTrXL) architecture, enabling the agent to maintain extended temporal contexts critical for complex sequential decision-making. We validate our approach on canonical control problems—the double integrator and Van der Pol oscillator—and demonstrate its effectiveness on multi-phase spacecraft trajectory optimization scenarios. Results show near-optimal performance comparable to analytical solutions where available, confirming the transformer’s ability to process long-range dependencies offers particular advantages for autonomous mission planning across dynamically distinct phases.

Dr. Amit Jain:  Dr. Amit Jain is a postdoctoral associate in the Department of Aeronautics and Astronautics at MIT, where he focuses on applying artificial intelligence and machine learning techniques to aerospace controls. He earned his master’s and Ph.D. in Aerospace Engineering from Pennsylvania State University.