4 AI MAGAZINE
Of course, deploying large numbers of integrated
intelligent agents, each utilizing multiple AI tech-
nologies, is the end vision of space AI technologists.
The first major step toward this vision was the remote
agent experiment (RAX) (Muscettola et al. 1998,
Bernard et al. 2000). RAX controlled the Deep Space
1 spacecraft for two periods totaling approximately
48 hours in 1999. RAX included three AI technolo-
gies: a deliberative, batch planner-scheduler, a robust
task executive, and a model-based mode identifica-
tion and reconfiguration system.
More recently, the autonomous sciencecraft has
controlled the Earth Observing- 1 mission for almost
10 years as this article goes to press. This run of operations includes more than 50,000 images acquired
and hundreds of thousands of operations goals. The
autonomous sciencecraft (Chien et al. 2005a)
includes three types of AI technologies: a model-based, deliberative, continuous planner-scheduler
(Tran et al. 2004, Rabideau et al. 2004), a robust task
executive, and onboard instrument data interpretation including support vector machine-learning
derived classifiers (Castano et al. 2006, Mandrake et
al. 2012) and sophisticated instrument data analysis
(see figure 1) (Thompson et al. 2013b).
Many individual AI technologies have also found
their way into operational use. Flight operations such
as science observation activities, navigation, and
communication must be planned well in advance.
AI-based automated planning has found a natural
role to manage these highly constrained, complex
operations. Early successes in this area include the
ground processing scheduling system (Deale et al.
1994) for NASA space shuttle refurbishment and the
SPIKE system used to schedule Hubble Space Telescope operations (Johnston and Miller 1994). SPIKE
enabled a 30 percent increase in observation utilization (Johnston et al. 1993) for Hubble, a major
impact for a multibillion dollar mission. Also impressive is that SPIKE or components of SPIKE have been
or are being used for the FUSE, Chandra, Subaru, and
Spitzer missions. AI-based planning and scheduling
are also in use on the European Space Agency’s Mars
express and other missions. For a more complete survey of automated planning and scheduling for space
missions see Chien et al. (2012a).
In this issue, the article by Mark D. Johnston,
Daniel Tran, Belinda Arroyo, Sugi Sorensen, Peter Tay,
Butch Carruth, Adam Coffman, and Mike Wallace
describes the deep space network ( DSN ) scheduling
engine (DSE) component of a new scheduling system
that provides core automation functionality for
scheduling of NASA’s deep space network, supporting
scores of missions with hundreds of tracks every
week. The article by Russell Knight, Caroline
Chouinard, Grailing Jones, and Daniel Tran describes
the application and adaptation of the ASPEN (
automated scheduling and planning environment)
framework for operations of the Orbital Express (OE)
Because space missions produce enormous petas-cale data sets, machine learning, data analysis, and
event recognition for science and engineering purposes has been another fertile area for application of
AI techniques to space applications (Fayyad, Haussler, and Stolorz 1996). An early success was the use
2050 2100 2150 2200 2250 2300 2350
Abundance Map Generated Onboard EO- 1
During November 2011 Over;ight
Abundance Map Generated by Expert
Human Interpretation (Kruse et al. 2003)
Figure 1. Onboard Spectral Analysis of Imaging Spectroscopy Data During 2011–2012 Demonstrated on EO- 1.
Onboard software performed spectral endmember detection and mapping, enabling automatic abundance mapping to reduce data volume
by orders of magnitude (Thompson et al 2013). These onboard automatically derived compositional maps (at left) were consistent with
prior expert human interpretations (at right).