Computational Challenges for Model-based Autonomous Systems Prof. Brian C. Williams Massachusetts Institute of Technology Artificial Intelligence & Space Systems Laboratories Space Systems & Artificial Intelligence Laboratories Massachusetts Institute of Technology
Polar Lander Leading Diagnosis: • Legs deployed during descent. • Noise spike on leg sensors latched by software monitors. • Laser altimeter registers 50ft. • Begins polling leg monitors to determine touch down. • Latched noise spike read as touchdown. Idea: Support programmers with • Engine shutdown at ~50ft. embedded languages that avoid commonsense mistakes, by reasoning from hardware models. Reactive Model-based Programming Space Systems & Artificial Intelligence Laboratories Massachusetts Institute of Technology
Model-based Executives automate all reasoning about system interactions. •Scheduling •Command confirmation •Diagnosis •Commanding •Configuration •Repair . . .
Model-based Programs Interact Directly with State Model-based programs Embedded programs interact with interact with plant state: plant sensors and actuators: • Read state • Read sensors • Write state • Set actuators Model-based Embedded Program Embedded Program Obs Cntrl S’ Model-based Executive Obs Cntrl S S Plant Plant Programmers must map Model-based executives map between states and automatically between states and sensors/actuators. sensors/actuators. Space Systems & Artificial Intelligence Laboratories Massachusetts Institute of Technology
Example: The model-based program sets the state to thrusting, and the model-based executive . . . . Oxidizer tank Fuel tank Oxidizer tank Fuel tank Deduces that Plans actions thrust is off, and to open the engine is healthy six valves Deduces that a valve failed - stuck closed Determines that valves on the backup engine will achieve thrust, and plans needed actions. Space Systems & Artificial Intelligence Laboratories Massachusetts Institute of Technology
Modeling Complex Behaviors through Probabilistic Concurrent Constraint Automata Valve Valve Stuck Stuck 0.01 0.01 Open Open open open 0. 01 0. 01 Open Open Close Close 0. 01 0. 01 Stuck Stuck Closed Closed closed 0.01 closed 0.01 inflow = outflow = 0 • Complex, discrete behaviors • modeled through concurrency, hierarchy and non-determinism. • Anomalies and uncertainty • modeled by probabilistic transitions • Physical interactions • modeled by discrete and continuous constraints • Timing • modeled by simple temporal networks
Model-based Executive Model-based Autonomy (Livingstone, Titan, Kirk..) Architecture Control Program Sequencer Model-based Program Searches for optimal Control Program Searches for optimal feasible threads of execution feasible threads of execution � Executes concurrently � Preempts Plan Plans � Asserts and queries states Failures � Chooses based on reward Performs lazy scheduling Performs lazy scheduling � Expresses temporal and resource constraints State estimates State goals System Model Mode Estimation Mode Reconfiguration Valve Valve Stuck Stuck 0.01 0.01 Tracks Open Open open open Finds 0. 01 0. 01 Plans likely Open Open Close Close best 0. 01 0. 01 Stuck Stuck reactively state Closed Closed closed closed target 0.01 0.01 inflow = outflow = 0 trajectories Computational Challenges: Deductive Controller • Propositional Satisfiability • Optimal CSPs Plant • Graph-based Planning Observations Commands • Scheduling
Future autonomy requires reasoning about hybrid discrete/continuous systems Detecting subtle failures Coordinating fleets of agile vehicles − ≥ − x x d Mb p q pq 1 − ≥ − − ≥ and x x x x d d Mb new q p pq 2 p q old estimate: estimate: − − ≥ ≥ − and y or x y x d d Mb p q q p pq 3 d − − ≥ ≥ − and y y d Mb or y y d q p p q pq 4 − ≥ or y y 4 d continuous state Y q p ∑ Probabilistic ≤ and b 3 evolution within mode X pqk mode transition = k 1 Compute hybrid of: Compute hybrid of: • temporal constraint problem • HMM belief update • mixed integer linear program • Kalman filtering Space Systems & Artificial Intelligence Laboratories Massachusetts Institute of Technology
Recommend
More recommend
