Editor: Richard Doyle Jet Propulsion Lab rdoyle@jpl.nasa.gov AI in Space to science return. This is the main role of the Mars-Express Scheduling Architecture (Mexar2), an AI-based tool in daily use on the Mars-Express mission since February 2005. Mexar2 supports space mission planners continuously as they plan data downlinks from the spacecraft to Earth. The tool lets planners work at a higher abstraction level while it performs low-level, often-repetitive tasks. It also helps them produce a plan rapidly, explore alternative solutions, and choose the most robust plan for execution. Addition- ally, planners can analyze any problems over multiple days and identify payload overcommitments that cause resource bottlenecks and increase the risk of data losses. Mexar2 has significantly increased the data return over the whole Mars-Express mission duration. It’s effectively become a work companion for mission planners at the Euro- pean Space Agency’s European Space Operations Center (ESOC) in Darmstadt, Germany. Dumping data from Mars Launched on 2 June 2003, Mars-Express is the ESA’s first mission to orbit another planet. Notwithstanding a limited budget, the mission has ambitious goals for its sci- entific experiments. The orbiter is equipped with seven payloads, each gathering novel planetary information. The “Mars-Express Mission” sidebar describes some mission challenges relevant to transmitting the data to Earth. In addition to the large amount of science data pro- duced through its payload activities, the spacecraft’s onboard platform produces housekeeping data through its health-monitoring and verification tasks. All these data are intended for transfer to Earth during bounded downlink sessions. Mars-Express has a single-pointing system—that is, dur- ing regular operations, the space probe points either at Mars to perform payload operations or at Earth to download the produced data. Consequently, the mission is designed to first store data in the onboard memory and then transfer it to Earth. The main issue in this process is to decide sequences of spacecraft operations—dump plans—to deliver the on- board memory contents during the available downlink win- dows. The mission produces about 2 to 3 Gbits of data each day, but the downlink channel transmits between 28 and 182 Kbits per second—a rate that can be insufficient for communicating all the data produced. As figure 1 shows, the onboard memory is subdivided into different banks, or packet stores, for both science and housekeeping data. Mars-Express manages each science store cyclically, so if new data is produced before the pre- vious data is dumped, the older data gets overwritten. This means the related observation experiments must be rescheduled. The onboard memory reserved for science data is about 9 Gbits. However, some payloads produce very large data files—for example, the High-Resolution Stereo Camera (HRSC) produces files up to 1 Gbit per observation—so usage often comes close to the packet store capacities. Ir- regularly distributed transmission windows and different transmission rates also contribute to this problem. Finally, some mission instruments use different compression algo- rithms, adding another uncertainty factor to data production. The Mars-Express Mission Planning Group computes nominal dump plans automatically from expected pay- load activity. However, when mission planners analyze their housekeeping data, they not infrequently discover that onboard data exceeds expectations. This means they must recompute the dump plan that implements the mem- ory download policy. The overall problem is quite com- plex and, as explained later, contains different work constraints. Modeling the problem Solving the memory-dumping problem optimally is a dif- ficult combinatorial problem. Additionally, for deep-space missions, the solution entails a set of critical decisions that D eep-space missions carry an ever larger set of differ- ent and complementary onboard payloads. Each payload generates data, and synthesizing it for optimized downlinking is one way to reduce the ratio of mission costs Mexar2: AI Solves Mission Planner Problems Amedeo Cesta, Gabriella Cortellessa, Simone Fratini, and Angelo Oddi, Italian National Research Council Michel Denis, Alessandro Donati, Nicola Policella, Erhard Rabenau, and Jonathan Schulster, European Space Agency 12 1541-1672/07/$25.00 © 2007 IEEE IEEE INTELLIGENT SYSTEMS Published by the IEEE Computer Society