Follow this link to skip to the main content
  NASA Logo
Jet Propulsion Laboratory
California Institute of Technology
+ View the NASA Portal
Search JPL
JPL Home Earth Solar System Stars and Galaxies Technology
JPL Robotics
Search Robotics
 
Home Page
Applications
Projects
Tasks
Groups
People
Systems
Facilities
Publications
Patents
Awards
News
Image Gallery
Video Gallery
Charter
Links
 
RESEARCH TASK
underline
Task Page Publications Image Gallery Videos
Risk-Aware Machine Learning for Resilient Space Exploration

Description:
The objective of the task is to develop a risk-aware, machine learning-based planning and decision making capabilities that will enable a spacecraft to dynamically adapt to its environment while explicitly reasoning about risk to ensure safety.

Space exploration is fundamentally about exploring the unknown. When designing spacecraft autonomy, the traditional approach is to develop a point design that works robustly for all expected situations. However, when future missions dare to explore worlds that are even more remote, the traditional approach will fail due to the large uncertainty. This challenge requires to an adaptive approach, where the autonomous spacecraft learns about the risks of its new environment, updates its internal models, and flexibly adjusts its behavior.

Take, for example, Curiosity’s unexpected slippage on a low-slope, sandy terrain. The prediction of slippage of the Mars rover is largely based on the model derived from extensive experiments on Earth. However, Curiosity has repeatedly observed excessive (up to 80%) slip on a low-slope sandy terrain, where the Earth-based model predicts only 10-20% of slip. Such a high slip posed a major risk on the mission since it may result in embedding of the rover. A slow driving speed of Curiosity allowed the ground operators to adapt to the unexpected situation. However, in a future missions with more agile and autonomous mobility and increased communication latency, on-board learning and adaptation will be essential.

While machine learning has proven useful in Earth satellite missions (e.g., scene classification), it is not considered to be ready for mission-critical applications, such as on-board activity planning, for two reasons: 1) traditional machine learning requires a significant volume of training data, while space missions require quick adaptation to new environment; 2) on-board learning may result in potentially risky behavior that cannot be checked a-priori by ground operation. Overcoming these challenges is critical for advancing the frontier of space exploration.

Point of Contact:  Masahiro Ono - Jet Propulsion Laboratory

Sponsored by:  President and Directors Fund



Privacy/Copyright Image Policy Glossary Sitemap Feedback Contact Us
  National Aeronautics and Space Administration website.