Learning-Based Methods for Spacecraft Dynamics Modeling, Filtering, and Predictive Control
Spacecraft motion under model uncertainty arises in many on-orbit assembly, servicing, and assistance scenarios, including tasks requiring manipulation of unknown grappled objects. Traditionally, adaptive model-based control approaches have relied on an analytical dynamics model with a set of parame...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2022
|
| Online Access: | https://hdl.handle.net/1721.1/144750 https://orcid.org/ 0000-0002-6090-6934 |
| _version_ | 1826206263330996224 |
|---|---|
| author | Parker, William E. |
| author2 | Linares, Richard |
| author_facet | Linares, Richard Parker, William E. |
| author_sort | Parker, William E. |
| collection | MIT |
| description | Spacecraft motion under model uncertainty arises in many on-orbit assembly, servicing, and assistance scenarios, including tasks requiring manipulation of unknown grappled objects. Traditionally, adaptive model-based control approaches have relied on an analytical dynamics model with a set of parameters that are estimated from observations of effective spacecraft dynamics. Without extensive a priori knowledge of the system under study, however, it can be difficult to identify a parametric model structure that accurately captures the dynamics of the system. In this work, the author proposes a new approach for learning unknown ``hard-to-model'' spacecraft dynamics in a non-parametric way using techniques including Gaussian process regression, deep evidential regression, and a novel particle filter regression scheme. These non-parametric and uncertainty-aware methods allow previously unmodeled dynamics to be learned onboard a spacecraft in real-time with very little a priori knowledge of the system required, but come with increased computational cost.
State estimation and control tasks typically rely on accurate process and observation models, but these process models have historically been analytical and parametric, requiring advance knowledge of the system. In this work, uncertainty-aware learned non-parametric dynamics models are used for state estimation filtering, model predictive control, and Fault Detection, Isolation, and Recovery (FDIR) scenarios. The Uncertainty-Aware Regression Unscented Kalman Filter (UAR-UKF) is developed and applied to perform state estimation for a nonlinear dynamical system using a learned process model. The Uncertainty-Aware Regression Bayesian Filter (UAR-BF) is also designed to capitalize on the learned process model's ability to perform state and covariance transitions, and uses Gaussian conflations instead of the Kalman gain to compute a posterior state and covariance at each timestep. A non-parametric model predictive control framework is also discussed, where optimal control trajectories are computed over a finite receding time horizon using a learned process model. An example scenario is presented to highlight the utility of learning-based methods for fault detection, isolation, and recovery after a sudden actuator failure. Learning-based methods are shown to be favorable compared to parametric modeling methods for both the filtering and control applications in simulation and on real robotic systems operating in microgravity on the International Space Station. The tools developed in this work are generally applicable and potentially useful for non-parametric learning and control of any complex, uncertain system in which little or no a priori knowledge is available. |
| first_indexed | 2024-09-23T13:26:40Z |
| format | Thesis |
| id | mit-1721.1/144750 |
| institution | Massachusetts Institute of Technology |
| last_indexed | 2024-09-23T13:26:40Z |
| publishDate | 2022 |
| publisher | Massachusetts Institute of Technology |
| record_format | dspace |
| spelling | mit-1721.1/1447502022-08-30T03:28:51Z Learning-Based Methods for Spacecraft Dynamics Modeling, Filtering, and Predictive Control Parker, William E. Linares, Richard Massachusetts Institute of Technology. Department of Aeronautics and Astronautics Spacecraft motion under model uncertainty arises in many on-orbit assembly, servicing, and assistance scenarios, including tasks requiring manipulation of unknown grappled objects. Traditionally, adaptive model-based control approaches have relied on an analytical dynamics model with a set of parameters that are estimated from observations of effective spacecraft dynamics. Without extensive a priori knowledge of the system under study, however, it can be difficult to identify a parametric model structure that accurately captures the dynamics of the system. In this work, the author proposes a new approach for learning unknown ``hard-to-model'' spacecraft dynamics in a non-parametric way using techniques including Gaussian process regression, deep evidential regression, and a novel particle filter regression scheme. These non-parametric and uncertainty-aware methods allow previously unmodeled dynamics to be learned onboard a spacecraft in real-time with very little a priori knowledge of the system required, but come with increased computational cost. State estimation and control tasks typically rely on accurate process and observation models, but these process models have historically been analytical and parametric, requiring advance knowledge of the system. In this work, uncertainty-aware learned non-parametric dynamics models are used for state estimation filtering, model predictive control, and Fault Detection, Isolation, and Recovery (FDIR) scenarios. The Uncertainty-Aware Regression Unscented Kalman Filter (UAR-UKF) is developed and applied to perform state estimation for a nonlinear dynamical system using a learned process model. The Uncertainty-Aware Regression Bayesian Filter (UAR-BF) is also designed to capitalize on the learned process model's ability to perform state and covariance transitions, and uses Gaussian conflations instead of the Kalman gain to compute a posterior state and covariance at each timestep. A non-parametric model predictive control framework is also discussed, where optimal control trajectories are computed over a finite receding time horizon using a learned process model. An example scenario is presented to highlight the utility of learning-based methods for fault detection, isolation, and recovery after a sudden actuator failure. Learning-based methods are shown to be favorable compared to parametric modeling methods for both the filtering and control applications in simulation and on real robotic systems operating in microgravity on the International Space Station. The tools developed in this work are generally applicable and potentially useful for non-parametric learning and control of any complex, uncertain system in which little or no a priori knowledge is available. S.M. 2022-08-29T16:09:08Z 2022-08-29T16:09:08Z 2022-05 2022-06-09T16:14:45.069Z Thesis https://hdl.handle.net/1721.1/144750 https://orcid.org/ 0000-0002-6090-6934 In Copyright - Educational Use Permitted Copyright MIT http://rightsstatements.org/page/InC-EDU/1.0/ application/pdf Massachusetts Institute of Technology |
| spellingShingle | Parker, William E. Learning-Based Methods for Spacecraft Dynamics Modeling, Filtering, and Predictive Control |
| title | Learning-Based Methods for Spacecraft Dynamics Modeling, Filtering, and Predictive Control |
| title_full | Learning-Based Methods for Spacecraft Dynamics Modeling, Filtering, and Predictive Control |
| title_fullStr | Learning-Based Methods for Spacecraft Dynamics Modeling, Filtering, and Predictive Control |
| title_full_unstemmed | Learning-Based Methods for Spacecraft Dynamics Modeling, Filtering, and Predictive Control |
| title_short | Learning-Based Methods for Spacecraft Dynamics Modeling, Filtering, and Predictive Control |
| title_sort | learning based methods for spacecraft dynamics modeling filtering and predictive control |
| url | https://hdl.handle.net/1721.1/144750 https://orcid.org/ 0000-0002-6090-6934 |
| work_keys_str_mv | AT parkerwilliame learningbasedmethodsforspacecraftdynamicsmodelingfilteringandpredictivecontrol |