Deep Learning Enabled Model Predictive Control for Optimized and Autonomous Science Data Collection
Abstract
Planetary Science missions stand to benefit from recent developments in miniaturization of satellite technology (e.g. CubeSats) and increased on-board autonomous operation. Whether utilizing onboard processes for autonomous instrument and spacecraft operations (such as descent or landing) or to controlling a network of Smallsats, balloons, or rovers - autonomous computing platforms can optimize both power utilization and the science data collection process. Planetary science missions have unique operational challenges and constraints - communication delays, weight limitations, processing and power limitations, and radiation effects. Through internal R&D and NASA funding, we have developed algorithms supporting optimized data collection from space using an architecture based upon Model Predictive Control (MPC). MPC is an ideal framework for autonomous, adaptive, real-time control of complex systems because it re-optimizes actions with respect to multiple weighted goals and constraints at every time step. Because of this proven capability, it is often the starting point for software control architecture design of many terrestrial applications including autonomous cars, robotic vision systems and it has been proposed for control of distributed spacecraft. However, MPC algorithms are known to be computationally challenging and a difficult framework to incorporate heritage data. This paper extends previous work by embedding aspects of a specific machine learning approach called deep learning (DL). DL algorithms, a neural network based black box mapping function, complements the MPC architecture and facilitates use of previous data bases for pattern recognition and scene classification. We show results of recent work targeting on-orbit autonomous lidar characterization of planetary features using DL algorithms to replace conventional pattern recognition algorithms and adaptively control the lidar beam targeting. The similar architecture approach can be extended to autonomously control flyby or landing on planetary bodies, and exploration of the surface and caves by swarms of autonomous platforms.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2018
- Bibcode:
- 2018AGUFM.P41D3763L
- Keywords:
-
- 1942 Machine learning;
- INFORMATICSDE: 5799 General or miscellaneous;
- PLANETARY SCIENCES: FLUID PLANETSDE: 6299 General or miscellaneous;
- PLANETARY SCIENCES: SOLAR SYSTEM OBJECTSDE: 5464 Remote sensing;
- PLANETARY SCIENCES: SOLID SURFACE PLANETS