The challenge of space exploration has dramatically shifted from simple fly-bys to micro-rovers that can alight upon several asteroids, collect the most interesting geological samples, and return with their findings. This challenge will not be answered through billion dollar missions with 100 member ground teams, but through innovation. Future space exploration will be enabled in significant part by space explorers that are self-reliant and capable of handling unexpected situations; they must balance curiosity with caution. Self-reliance of this sort can only be achieved through an explicit understanding of mission goals and the ability to reason from a model of how the explorer and its environment can support or circumvent these goals. Robustness of this sort can only be achieved by careful coordination of the complex network of sensors and actuators within a spacecraft. Given the complexity of current (and future) spacecraft, such fine-tuned coordination is ordinarily a nearly-impossible task, both conceptually and as a software engineering undertaking. Such coordination is also essential to creating and operating future networked embedded systems, such as networks of cooperative air and ground vehicles that perform search and rescue. Similar levels of robustness and ease of use are equally relevant in more down to earth contexts within complex embedded systems of the sort found in automobiles and naval ships.

Our research confronts these challenges by introducing a new automated reasoning paradigm called model-based autonomy. We envision model-based explorers that are programmed rapidly and simply by specifying strategic guidance in the form of a few high-level control behaviors, called model-based programs. These control programs, along with a commonsense model of its hardware and its environment, enable an explorer to control and monitor its hidden state according to the strategic guidance. To respond correctly in novel, time-critical situations, our explorers use their onboard models to perform extensive commonsense reasoning within the reactive control loop, something that conventional AI wisdom had suggested was not feasible.   Elevated to a higher level, model-based programs will enable a single human operator to coach a team of agile search and rescue vehicles or a colony of robotic astronauts.

The following subsections describe our recent research areas:

• Model-based Programming of Fault Aware Systems

• Coaching Robotic Networks Through Model-based Programming

• Hybrid Model-based Adaptive Systems

Our recent research efforts have focused on creating Mars Exploration Rovers that are fault aware, systems that coach air vehicle teams that themselves can cooperate intelligently, and walking robots that recover from stumbles with great agility.  We have achieved significant advances in real-time reasoning along four fronts: model-based diagnosis and estimation, model-based planning and execution, knowledge compilation, and optimal deductive reasoning. These capabilities were applied to simulation scenarios for three space missions (the NASA Mars Exploration Rover and the Mars Science Laboratory Missions, and MIT Spheres), a Martian habitat, a humanoid robot and a team of fixed wing air vehicles.