Deliberative & Hybrid Control ROBOTICS COE 584 Deliberative & Hybrid Control

Lecture Outline Deliberative control Hybrid control Types of layer organization selection advising adaptation postponement Examples of hybrid control AuRA, Atlantis SSS, PRS

Deliberative Systems Purely deliberative systems are considered the classical control architecture, since they were the first to be tried In AI, classical deliberative, planner-based architectures were used for reasoning about actions in various non-physical domains, such as chess As a result, the same architectures were applied to robotics as well

In the 1960’s: Shakey In the late 1960's, the state-of-the-art in machine vision was used to process visual information on a robot called Shakey, the forerunner of many AI-inspired robotics projects. Shakey used a classical planner as the underlying structure to decide what to do. What is planning?

Planning as Search Planning is looking ahead, searching The goal is a state The robot's entire state space is enumerated, and searched, from the current state to the goal state Different paths are tried until one is found that reaches the goal If the optimal path is desired, then all possible paths must be considered in order to find the best one

SPA = Planner-based Planner-based (deliberative) architectures typically involve three generic sequential steps or functional modules: 1) sensing (S) 2) planning (P) 3) acting (A), executing the plan Thus, they are called SPA architectures SPA has serious drawbacks What are they?

Problem 1: Time Complex state spaces: Dynamic worlds: very slow plan generation Dynamic worlds: out of date plans (latency)

Problem 2: Space Representation of state space may be very large Search tree (intermediate plan data) may be very large Modern machines have virtual memory (page to disk), but swapping is very slow

Problem 3: Representation Representation for planning has two parts: Knowing the state of the world Predicting the outcome of actions State representation assumed to be: complete accurate current predictable

Problem 3: Representation Sensors have: noise inaccuracies aliasing (partial observability) Effectors are: unpredictable unreliable None of the assumptions are valid!

Problem 4: Execution Execution is assumed to be: But: sequential reliable unique (one actor) But: blind execution of long sequences of unreliable actions will fail E.g., p(success | 1 action) = 0.90 => p(success | 10 actions) = 0.35

Deliberative Summary In short, deliberative (SPA) approaches: require search (slow) require representations (hard) encourage open-loop execution (dangerous)

Opposition to SPA As a consequence, much opposition from real robot practitioners mounted against SPA architectures In the early/mid 1980's alternatives were proposed reactive systems hybrid systems What happened to purely deliberative systems?

Role of Pure Deliberation Pure deliberation is alive and well in other domains, like game playing (chess, go, etc.) and other static worlds with plenty of time to plan

Planners Live On in Robotics The SPA approach has not been abandoned, it has been expanded Given the two fundamental problems with purely deliberative approaches, we can augment them: search/planning is slow, so save/cache important and/or urgent decisions; open-loop plan execution is bad, use closed-loop feedback, and be ready to respond or re-plan when the plan fails.

Reusing Plans Some frequently useful planned decisions may need to be reused, so to avoid planning, an intermediate layer may cache and look those up These can be intermediate-level actions (ILAs) macro operators: plans compiled into more general operators for future use

Universal Plans Suppose for a given problem, all possible plans are generated for all possible situations in advance, and stored If for each situation a robot has a pre-existing optimal plan, it can react optimally, be reactive and optimal It has a universal plan (These are complete reactive mappings)

Viability of Universal Plans A system with a universal plan is reactive; the planning is done at compile-time, not at run-time Universal plans are not viable in most domains, because they require that: the world must be deterministic the world must not change the goals must not change The world is too complex (state space is too large)

Situated Automata A formal notion of finite state machines whose inputs are connected to sensors and whose outputs are connected to effectors are called situated automata. Situated means existing in and interacting with a complex world, and automata is the formal name for FSMs (formally: finite state automata). Situated automata are used to create reactive principled control systems.

Control w/ Situated Automata Situated automata can be constructed in two basic ways: By hand (i.e., the designer puts FSMs together), as in the Subsumption Architecture). By pre-compiling a complete plan (similar to Universal Plans, but reduced down to circuits of FSMs). This requires the use of a special programming language that implements the right semantics and compiles down into FSM circuitry, as Rex and Gapps.

Domain Knowledge A key advantage of pre-compiled systems is that domain knowledge, i.e., information that the designer has about the environment, the robot, and the task, can be embedded into the system in a principled way Then, the system is compiled into a reactive circuit, so the knowledge does not have to be reasoned about (or planned with) explicitly, in real-time

Disadvantages A key disadvantage of pre-compiled systems is that it quickly becomes prohibitively large to enumerate the state space of a real robot, and thus pre-compiling generally does not scale up to complex systems Another disadvantage is common to compiled or hard-wired systems: the result is not flexible in the presence of changing environments, tasks or goals

Inventing Hybrid Control The basic idea is simple: we want the best of both worlds (if possible) The goal is to combine closed-loop and open-loop execution That means to combine reactive and deliberative control This implies combining the different time-scales and representations This mix is called hybrid control

Organizing Hybrid Systems A hybrid system typically consists of three components: a reactive layer a planner a layer that puts the two together Hybrid architectures are often called three-layer architectures (TLA) The planner and the reactive system are both standard, as we have covered them so far

The Magic Middle The middle layer has a hard job: 1) compensate for the limitations of both the planner and the reactive system 2) reconcile their different time-scales 3) deal with their different representations 4) reconcile any contradictory commands between the two This is the challenge of hybrid systems

Interaction of Layers Hierarchical integration Planning guides reaction Coupled planning & reacting

Dynamic Re-planning Reaction can influence planning Any "important" changes discovered by the low-level controller are passed back to the planner in a way that the planner can use to re-plan The planner is interrupted when even a partial answer is needed in real-time The reactive controller (and thus the robot) is stopped if it must wait for the planner to tell it where to go.

Planner-Driven Reaction Planning can influence reaction Any "important" optimizations the planner discovers are passed down to the reactive controller The planner’s suggestions are used if they are possible and safe Who has priority, planner or reactor?

Types of Interaction Selection: Planning is viewed as configuration Advising: Planning is viewed as advice giving Adaptation: Planning is viewed as adaptation of controller Postponing: Planning is viewed as a least commitment process

Selection Example: AuRA R. Arkin (1986) Planning is viewed as configuration Initial A* planner integrated with schema-based controller Provides modularity, flexibility, and adaptability

AuRA Schematic

Advising Example: Atlantis E. Gat (1991) (JPL) Three layers: controller, sequencer, deliberator Asynchronous, heterogeneous: reactivity and deliberation Implemented in ALFA (A Language for Action) Planning as advice giving, not decree Notion of cognizant failure Tested on NASA rovers Rocky 4

Atlantis Schematic

Adaptation Example: Planner-Reactor D. Lyons (1992) Continuous modification of a reactive control system Planning is a form of reactor adaptation Adaptation is on-line rather than off-line deliberation Planning is used to remove performance errors when they occur Uses a particular underlying mathematical model called a process algebra Tested in both assembly cell and grasp planning

Planner-Reactor Architecture GOALS PLANNER ADAPTION REACTOR ACTION WORLD REACTIONS PERCEPTIONS PERCEPTION SENSING

Postponing Example: PRS PRS = Procedural Reasoning System Georgeff and A. Lansky (1987) Least commitment via plan elaboration postponement Tested on SRI Flakey

Flakey the robot

PRS Schematic

Another Example: SSS J. Connell (1992) SSS = Servo Subsumption Symbolic 3 layers: servo, subsumption, symbolic World models are a convenience, not a necessity Symbolic: where-to-next (discrete time) Subsumption: where-to-go-now Servo: making it go (continuous time) Tested on TJ

SSS Implementation: T J

More Examples Multi-valued logic SOMASS hybrid assembly system C. Malcolm and T. Smithers (Edinburgh U.) cognitive/subcognitive components planning as configuration Agent architecture B. Hayes-Roth (Stanford) physical and cognitive levels functional boundary blurry Multi-valued logic Saffiotti, Konolige, Ruspini (SRI)

Even More Examples Supervenience Teleo-reactive agent architecture L. Spector (1992, U. of Maryland) Multiple levels of abstraction Teleo-reactive agent architecture Benson and N. Nilsson (1995, Stanford) Planning yields TR operator tree Reactive Deliberation M. Sahota (1993, U. of British Columbia) Robosoccer

Still More Examples Theoagent Generic Robot Architecture T. Mitchell (CMU, 1990) Reacts when it can plans when it must Emphasis on learning Generic Robot Architecture Noreils and Chatila (1995, France) 3 levels: planning, control system, functional Dynamical Systems Approach Schoner and Dose (1992) Planning is selecting and parameterizing behavioral fields Behaviors use vector summation

And Still More Examples Integrated path planning and dynamic steering control Krogh and C. Thorpe (1986, CMU) Relaxation over grid-based model with potential fields controller Planner generated waypoints for controller Many others (including several for UUVs)

Hybrids Everywhere? Hybrid systems are the most popular alternative for single-robot control Behavior-based systems are not used by quite as many researchers, but have more specialized niches (e.g., multi-robot systems) and more practical applications

Textbook Readings MM 13, 15