AI Planning Applications

1. AI Planning Applications Plan Execution Practical Applications of AI Planners

2. AI Planning Applications • Planning for Execution • Deep Space 1 and Remote Agent Experiment • Practical Applications of AI Planners • Common Themes

3. Deep Space 1 – 1998-2001 Photos: NASA

4. DS 1 at Comet Borrelly Photos: NASA

5. DS1 Domain Requirements Achieve diverse goals on real spacecraft • High Reliability • single point failures • multiple sequential failures • Tight resource constraints • resource contention • conflicting goals • Hard-time deadlines • Limited Observability • Concurrent Activity

6. DS1 Remote Agent Approach • Constraint-based planning and scheduling • supports goal achievement, resource constraints, deadlines, concurrency • Robust multi-threaded execution • supports reliability, concurrency, deadlines • Model-based fault diagnosis and reconfiguration • supports limited observability, reliability, concurrency • Real-time control and monitoring

7. DS1 Levels of Autonomy Listed from least to most autonomous mode: • single low-level real-time command execution • time-stamped command sequence execution • single goal achievement with auto-recovery • model-based state estimation & error detection • scripted plan with dynamic task decomposition • on-board back-to-back plan generation, execution, & plan recovery

8. DS 1 Levels of Autonomy

9. DS 1 Systems Planning Execution Monitoring

10. DS1 RAX Functionality Planner Scheduler/Mission manager (PS/MM) • generate plans on-board the spacecraft • reject low-priority unachievable goals • replan following a simulated failure • enable modification of mission goals from ground Executor (EXEC) • provide a low-level commanding interface • initiate on-board planning • execute plans generated both on-board and on the ground • recognize and respond to plan failure • maintain required properties in the face of failures Mode Indetification and Recovery (MIR) • confirm executive command execution • demonstrate model-based failure detection, isolation, and recovery • demonstrate ability to update on-board state via ground commands

11. DS1 Remote Agent (RA) Architecture

12. DS1 Planner Architecture

13. DS1 Diversity of Goals • Final state goals • “Turn off the camera once you are done using it” • Scheduled goals • “Communicate to Earth at pre-specified times” • Periodic goals • “Take asteroid pictures for navigation every 2 days for 2 hours” • Information-seeking goals • “Ask the on-board navigation system for the thrusting profile” • Continuous accumulation goals • “Accumulate thrust with a 90% duty cycle” • Default goals • “When you have nothing else to do, point HGA to Earth”

14. DS1 Diversity of Constraints • State/action constraints • “To take a picture, the camera must be on.” • Finite resources • power • True parallelism • the ACS loops must work in parallel with the IPS controller • Functional dependencies • “The duration of a turn depends on its source and destination.” • Continuously varying parameters • amount of accumulated thrust • Other software modules as specialized planners • on-board navigator

15. Temporal Constraints in DDL Command to EXEC in ESL DS1 Domain Description Language

16. DS1 Plan Fragment

17. DS1 RA Exec Status Tool

18. DS1 RA Ground Tools

19. DS1 – Flight Experiments17th – 21st 1999 • RAX was activated and controlled the spacecraft autonomously. Some issues and alarms did arise: • Divergence of model predicted values of state of Ion Propulsion System (IPS) and observed values – due to infrequency of real monitor updates. • EXEC deadlocked in use. Problem diagnosed and fix designed by not uploaded to DS1 for fears of safety of flight systems. • Condition had not appeared in thousands of ground tests indicating needs for formal verification methods for this type of safety/mission critical software. • Following other experiments, RAX was deemed to have achieved its aims and objectives.

20. DS 1 Experiment 2 Day Scenario

21. DS 1 SummaryObjectives and Capabilities

22. RAX Features • AI planner outer level with re-planning capability • Detailed constraint handling (e.g. time and resources) • Integration with system diagnostics and analysis • Integration with plan execution and monitoring • Rich knowledge modelling languages • Comprehensive user interfaces

23. ESA Spacecraft Planning Applications • APSI – Advanced Planning & Scheduling Initiative • AI, scheduling, constraint programming • MEXAR2 and RAXEM – Advanced Planning for Spacecraft Data Downlink and Telecommands Uplink • AI, mixed-initiative, scheduling, flow-network • SKeyP – SOHO Keyhole Planner • AI, scheduling, constraint programming • MrSPOCK – Mars Express Science Planning Opportunities Construction Kit • genetic algorithms, heuristic search

24. Earlier Spacecraft Planning Applications • Deviser • NASA Jet Propulsion Lab • Steven Vere, JPL • First NASA AI Planner • 1982-3 • Based on Tate’s Nonlin • Added Time Windows • Produced Voyager Mission Plans • Not used live, planned use for Uranus encounter

25. Earlier Spacecraft Planning Applications • Edinburgh T-SCHED Planner • Brian Drabble, AIAI, University of Edinburgh • British National Space Centre T-SAT Project • 1989 • Ground-based plan generation • 24 hour plan uploaded and executed live onboard UoSAT-II

26. PlanERS-1 and Optimum-AIV Photos: ESA, ArianeSpace

27. Photo: NASA

28. AI Planning Applications • Planning for Execution • Deep Space 1 and Remote Agent Experiment • Practical Applications of AI Planers • Common Themes

29. Some Practical Applications of AI Planning • Nonlin electricity generation turbine overhaul • Deviser Voyager mission planning demonstration • SIPE – a planner that can organise a …. brewery • Optimum-AIV – Spacecraft Assembly, Integration & Test • O-Plan – various uses – see next slides • SHOP/SHOP2 – Bridge Baron, etc. • Deep Space 1 – RAX – to boldly go…

30. Practical AI Planners

31. Practical Applications of AI Planning – SIPE-2: System for Interactive Planning and Execution David Wilkins, AI Center SRI International

32. Practical Applications of AI Planning – SIPE-2 Technology • Supports interactive planning, allowing humans and the system to cooperate in mixed-initiative planning • Efficiently reasons about actions to generate a novel sequence of actions that responds precisely to the situation at hand • Supports the giving of advice to the planner • Plans hierarchically at different levels of abstraction • Is domain-independent (multiuse) • Replans during execution • Generates parallel plans (allowing multiple agents) • Posts constraints and reasons about resources • Interacts with humans through a powerful graphical interface

33. Practical Applications of AI Planning – SIPE-2 Applications • Air campaign planning • Military operations planning • Oil Spill Response (including an example plan) • Production line scheduling • Construction planning • Planning the actions of a mobile robot • A range of toy problems and puzzles, such as Missionaries and Cannibals.

34. Practical Applications of AI Planning – SHOP2 Technology • Like its predecessor SHOP, SHOP2 generates the steps of each plan in the same order that those steps will later be executed, so it knows the current state at each step of the planning process. This reduces the complexity of reasoning by eliminating a great deal of uncertainty about the world, thereby making it easy to incorporate substantial expressive power into the planning system. • Like SHOP, SHOP2 can do axiomatic inference, mixed symbolic/numeric computations, and calls to external programs. • SHOP2 also has capabilities that go significantly beyond those of SHOP: • SHOP2 allows tasks and subtasks to be partially ordered; thus plans may interleave subtasks from different tasks. This often makes it possible to specify domain knowledge in a more intuitive manner than was possible in SHOP. • SHOP2 incorporates many features from PDDL, such as quantifiers and conditional effects. • If there are alternative ways to satisfy a method’s precondition, SHOP2 can sort the alternatives according to a criterion specified in the definition of the method. This gives a convenient way for the author of a planning domain to tell SHOP2 which parts of the search space to explore first. In principle, such a technique could be used with any planner that plans forward from the initial state. • So that SHOP2 can handle temporal planning domains, we have a way to translate temporal PDDL operators into SHOP2 operators that maintain bookkeeping information for multiple timelines within the current state. In principle, this technique could be used with any non-temporal planner that has sufficient expressive power.

35. Bridge Baron (Commercial Product) Practical Applications of AI Planning – SHOP2 Applications • Evacuation Planning • Evaluating Terrorist Threats • Fighting Forest Fires • Controlling Multiple UAVs • Software Systems Integration • Automated Composition of Web Services • Business Workflow Management • Project Planning • Creation of Virtual Educational Courses from Component Courses

36. Practical Applications of AI Planning – O-Plan Applications O-Plan has been used in a variety of realistic applications: • Construction Planning (Currie and Tate, 1991 and others) • Search & Rescue Coordination (Kingston et al., 1996) • Spacecraft Mission Planning (Drabble et al., 1997) • Engineering Tasks (Tate, 1997) • US Army Hostage Rescue (Tate et al., 2000a) • Noncombatant Evacuation Operations (Tate, et al., 2000b) • Biological Pathway Discovery (Khan et al., 2003) • Unmanned Autonomous Vehicle Command and Control • Web Services Composition and Workflow Management • O-Plan’s design was also used as the basis for Optimum-AIV (Arup et al., 1994), a deployed system used for assembly, integration and verification in preparation of the payload bay for flights of the European Space Agency Ariane IV launcher.

37. O-Plan Features A wide variety of AI planning features are included in O-Plan: • Domain knowledge elicitation • Rich plan representation and use • Hierarchical Task Network Planning • Detailed constraint management • Goal structure-based plan monitoring • Dynamic issue handling • Plan repair and re-planning in low and high tempo situations • Interfaces for users with different roles • Management of planning and execution workflow

38. AI Planning Applications • Planning for Execution • Deep Space 1 and Remote Agent Experiment • Practical Applications of AI Planners • Common Themes

39. Common Features for Practical AI Planners • Outer HTN “human-relatable” approach • Underlying detailed constraint handling (e.g. time and resources) • Integration with simulation and analysis • Integration with plan execution and monitoring • Rich knowledge modelling languages • Comprehensive user interfaces

40. Readings Deep Space 1 Papers • Bernard, D.E., Dorais, G.A., Fry, C., Gamble Jr., E.B., Kanfesky, B., Kurien, J., Millar, W., Muscettola, N., Nayak, P.P., Pell, B., Rajan, K., Rouquette, N., Smith, B., and Williams, B.C. Design of the Remote Agent experiment for spacecraft autonomy. Procs. of the IEEEAerospace Conf., Snowmass, CO, 1998. • Ghallab, M., Nau, D. and Traverso, P., Automated Planning – Theory and Practice, chapter 19, Elsevier/Morgan Kaufmann, 2004. Other Practical Planners • Tate, A. and Dalton, J. (2003) O-Plan: a Common Lisp Planning Web Service, invited paper, in Proceedings of the International Lisp Conference 2003, October 12-25, 2003, New York, NY, USA, October 12-15, 2003. • Ghallab, M., Nau, D. and Traverso, P., Automated Planning – Theory and Practice, chapters 22 and 23. Elsevier/Morgan Kaufmann, 2004

41. AI Planning Applications - Summary • Planning for Execution • Deep Space 1 and Remote Agent Experiment • Practical Applications of AI Planners • Common Themes

42. NASA Spacecraft Planning Applications • TBA