Pleased to share that our group has begun an AFOSR-supported project: “Online Robust Trajectory Planning of High-Dimensional Space Manipulators with Real-Time Optimization.”


We aim to make space manipulators more dependable by planning trajectories online, remaining resilient to uncertain targets, and optimizing in real-time.

Our system links global planning for the big picture, local planning for fast adjustments, and robust optimization to keep operations safe.

We will share progress on algorithm design, analysis, and large-scale simulations. 


1.      Zhizhuo Zhang*, Xiaoli Bai§. “Imitation Learning for Satellite Attitude Control under Unknown Perturbations”, 2025 AAS/AIAA Astrodynamics Specialist Conference.
 2.      Jacen Nisbet*, Xiaoli Bai§. “Effects of a New Global Thermospheric Density Prediction Model on Satellite Orbit Propagation”, 2025 AAS/AIAA Astrodynamics Specialist Conference.
 3.      Ruochen Wang*, Xiaoli Bai§. “SHAP analysis for a global thermospheric density prediction model”, 2025 AAS/AIAA Astrodynamics Specialist Conference.
 4.      Basil Khan*, Xiaoli Bai§. “Shape Reconstruction and Pose Estimation of an Unknown, Noncooperative, Rotating Target Using Structure-From-Motion”, 2025 AAS/AIAA Astrodynamics Specialist Conference.
 5.      Rabiul Hasan Kabir, Xiaoli Bai§. “Physics-ML fusion algorithm for estimation and prediction of relative motion of an unknown, non-cooperative, and tumbling space object”, 2025 AAS/AIAA Astrodynamics Specialist Conference.


This study addresses the motion and inertia parameter estimation problem of a torque-free, tumbling, non-cooperative space object (target) under long-term occlusions. 

To solve this problem, we employ a data-driven Gaussian process (GP) to simulate sensor measurements. In particular, we implement the multi-output GP to predict the projection measurements of a stereo-camera system onboard a chaser spacecraft. A product kernel, consisting of two periodic kernels, is used in the GP models to capture the periodic trends from non-periodic projection data. The initial guesses for the periodicity hyper-parameters of the GP models are intelligently derived from fast Fourier transform (FFT) analysis of the projection data.

Additionally, we propose an unscented Kalman filter–Gaussian process (UKF-GP) fusion algorithm for target motion and inertia parameter estimation. The predicted projections from the GP models and their derivatives are used as the pseudo-measurements for UKF-GP during long-term occlusion. Results from Monte Carlo (MC) simulations demonstrate that, for varying tumbling frequencies, the UKF-GP can accurately estimate the target’s motion variables over hundreds of seconds, a capability the conventional UKF algorithm lacks. 

See details: https://www.mdpi.com/1424-8220/25/3/647


This paper introduces a global thermospheric density prediction framework based on a deep evidential method. The proposed framework predicts thermospheric density at the required time and geographic position with given geomagnetic and solar indices. It is called global, which differentiates it from existing research that only predicts density along a satellite orbit. Through the deep evidential method, we assimilate data from various sources including solar and geomagnetic conditions, accelerometer-derived density data, and empirical models including the Jacchia-Bowman model (JB-2008) and the Naval Research Laboratory Mass Spectrometer and Incoherent Scatter Radar Extended (NRLMSISE-00) model. The framework is investigated on five test cases along various satellites from 2003 to 2015 involving geomagnetic storms with Disturbance Storm Time (Dst) values smaller than −50nT. 


Results show that the proposed framework can generate density with higher accuracy than the two empirical models. It can also obtain reliable uncertainty estimations. Global density estimations at altitudes from 200 to 550 km are also presented and compared with empirical models on both quiet and storm conditions.

See details: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2024SW004070


We are thrilled to announce that our proposal, DURIP: Testbed for Autonomous Proximity and Rendezvous Operations (ASPRO) with General Space Objects, has been selected! You can find the official announcement here: DOD Awards $43 Million to U.S. Universities

My vision for ASPRO is threefold: to serve as a cutting-edge testbed for our space operations research, to strengthen and establish new collaborations, and to create an engaging environment where students of all ages—even young children—can visit, gain hands-on experience, and be inspired in the exciting field of space exploration.


This paper considers the scenario where the chaser, carrying a monocular camera and a single-beam LIDAR sensor, closely inspects the target to collect high-resolution data. The chaser cannot change its trajectories, and the sensor can only observe the target partially due to its limited range and FOV. Unlike many previous works, we assume the same features cannot be captured at every time step, which is more realistic but challenging. 

A particular challenge for this work is how to design the reward function. In this work, rewards are assigned to the chaser states from a combined model with two components: feature detection score and sinusoidal reward. Estimated feature locations are required to calculate the sinusoidal reward, which are predicted by Gaussian Process models. Another Gaussian Process model provides the reward distribution, which the Bayesian Optimization then uses to determine the camera directional angles.

Simulations are conducted in both 2D and 3D domains. The results demonstrate that the proposed SOCRAFT algorithm can generally detect the maximum number of features within the limited camera range and field of view. 

https://www.mdpi.com/1424-8220/24/15/4831


Excited to announce our new project, "Machine Learning-Enabled Data Assimilation for Understanding and Forecasting the Variation of Thermospheric Neutral Density," funded by NASA’s Space Weather Science Applications Research-to-Operations-to-Research program! Understanding the dynamics of the Earth's thermosphere is crucial for various applications, from space weather forecasting to satellite orbital predictions and conjunction analysis. Our project aims to develop cutting-edge forecasting models that accurately predict thermospheric neutral density variations, even under challenging storm conditions. By leveraging machine learning techniques, we'll uncover the complex relationship between solar emissions and neutral density variations, providing valuable insights into space weather phenomena. Our ultimate goal? To provide a game-changing 3-day forecast of thermospheric density with unprecedented accuracy and quantifiable uncertainty, surpassing current methods. This project builds upon our ongoing NSF-supported research, with a focus on transitioning to operational use. I look forward to collaborating with my esteemed colleagues at NJIT and UC Boulder/NOAA on this groundbreaking endeavor. Stay tuned for updates!

We have been honored with a supplementary grant from the NSF in response to the call of NSF 21-577 Grand Challenges in Integrative Geospace Sciences: Advancing National Space Weather Expertise and Research toward Societal Resilience. We are deeply grateful for this expression of trust and are eager to embark on this new chapter. With this support, we look forward to working closely with our students and enhancing our community outreach activities. 

I am glad to share our latest published article that reflects our ongoing pursuit of accurate and reliable thermospheric neutral density prediction. In this work, we propose a deep evidential model-based framework. This innovative approach incorporates empirical models, accelerometer-inferred density from the CHAMP satellite, and geomagnetic and solar indices. Our results demonstrate the model's ability to predict thermospheric density with remarkable accuracy and reliable uncertainty during both quiet and storm times.

 Notably, the proposed model outperforms the Gaussian  Processes model used in our previous studies. Furthermore, it provides insightful aleatoric and epistemic uncertainties, which significantly contribute to the body of knowledge in this area.  

https://www.sciencedirect.com/science/article/pii/S0094576523003223

 

Introducing our latest publication, " Multi-step Bayesian Optimization-based intelligent task planning for an ozone observation satellite”, https://www.sciencedirect.com/science/article/pii/S0094576523001650

Dive into our recent research that extends the capabilities of Bayesian optimization (BO) for intelligent ozone observations. Traditional BO methods are limited by their myopic nature, which doesn't account for the impact of future decisions on the current action. In this study, we overcome this limitation by developing a novel Continuous Belief Tree Search algorithm, which enables a non-myopic, multi-step look-ahead optimization in continuous action domains.


 Our innovative approach incorporates a Control-effort based Upper Confidence Bound function as the acquisition function for BO, enhancing its effectiveness in tackling complex problems. Through comprehensive simulations, we demonstrate that our multi-step BO method substantially reduces model uncertainty when compared to the conventional single-step BO.
 

The basic assumption of the study is that we may not know what we shall look for until we see it. The Research Question is: Is there an intelligent observation strategy that is more informative and guided based on previous measurements than observing prior specified locations? Using measuring the total column ozone (TCO) as the application, we show that the proposed Bayesian optimization (BO)-based tasking method leads to both better prediction accuracy and better precision of the TCO distribution. https://arc.aiaa.org/doi/full/10.2514/1.A35249


In the situation of pollutant leakage from industrial facilities, comprehensive and accurate monitoring of the area of incidents is critical. In this paper, we present a Bayesian Optimization (BO)-based approach for such a task and use the Gaussian Processes (GPs) as the underlying surrogate model. Here is the link to the full-text access to a view-only version of our paper: https://link.springer.com/article/10.1007/s10846-022-01709-x


We are starting a new project "Intelligent Observing Non-cooperative Space Objects” funded by the “Dynamic Data and Information Processing” program at AFOSR. The goal of the project is to develop theoretical and computational methodologies that enable intelligent decisions from cooperative satellites on how to observe non-cooperative space objects. My interest in space robotics started in 2014 which led to my first Ph.D. student’s thesis and the interest from AFRL about our work. It takes a while for the work to be funded. Now we are grateful for the opportunity to focus on this work in the coming years. 


This paper is a culmination of our past years’ work on physics-based machine learning for orbit predictions. We have developed a strategy to integrate Gaussian Processes and particle filters. The methodology was tested on over 100 resident space objects in different orbit types. The prediction precision can be significantly improved for the majority of the cases. 

https://www.sciencedirect.com/science/article/pii/S0094576522003137


On this Earth Day, we are awarded an NSF grant, “ANSWERS: Prediction of Geoeffective Solar Eruptions, Geomagnetic Indices, and Thermospheric Density Using Machine Learning Methods.” I look forward to working with my wonderful team from Rutgers, NJIT, WVU, and MSU. More information about the project can be seen here: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2149747&HistoricalAwards=false

We are excited to be one of the seven selected teams: https://www.nsf.gov/news/special_reports/announcements/042622.jsp