Robot Control Architectures

Saphira architecture is a hybrid architecture which has no well-defined hierarchy or complexity layers, and is focused on mainly three features:

coherence deals with construction and maintainance of internal representations of the outwide world,
coordination of behaviors or determination of output of actions are accomplished using fuzzy logic and context dependent blending.
coordination with other agents using natural language or gestures.

Flakey of SRI is used as real robot and the task is Flakey is introduced to the office environment as a new employee, and then asked to perform delivery tasks. It may have a priori map and it should follow a human who introduces the office. It can additionally take advices on map-making or potential hazards. Some delivery task will be accomplished by planning the route and recovering errors when needed.
Coordination: It says that a layered abstraction approach makes complexity manageable but I think there are no contraints defined in the architecture to accomplish this.
Coherence: As complexity increases, agents must have a conception of their environment, in contrast to Connell, who says environment is the model, or others say world is its own representation.
Communication: A mobile agent will be of greater use if it interact effectively with other agents, understand task commands. Not a big step for this scheme.
Saphira architecture is an integrated sensing and control system, which includes Local Perception Space (LPS) at its center which contains different levels of representation (from occupancy grids, to geometric representation and to high level artifacts of the world). Internal artifacts (like it is a chair, doorway etc.) are viewed as beliefs of robot about the environment. Perception and action modules in levels of complexity, all interact with LPS. As another central module, Procedural Reasoning System (PRS) is used in more complex behaviors and in interaction with modules like speech input, schema library, topological planner.

At control level Saphira is behavior based, behaviors are written and combined using techniques based on fuzzy logic. These rules produce desirability function for each behavior, the fuzzy connectives are used to combine different behaviors based on their contexts (context depending blending), and defuzzification is used to choose preferred control among selected behaviors, generally taking average.
Basic behaviors take their inputs from the LPS and use information like occupancy data. More complex behaviors like goal-seeking behaviors (which are again basic behaviors), take input from artifacts. For example, the behavior cross-door uses the coordinates of a door artifact in the LPS as an input.
Basic behaviors are combined to form complex behaviors, where outputs of desirability functions for behaviors are combined (for example through a minimum operation as defined by context dependent blending and defuzzification). In such a way for example if there is an onbstacle ahead, and there are two choices, turn left or right, the selection is done according to the overall goal position.
This coordination mechanism is very similar to potential fields method, but two main advantages over it:

In context-depending blending, first behavioral preferences combined and then one summary control is chosen, but in potential fields, first preference summaries are combined and combined force is the combination of summaties. Potential fields is not mathematically formal, thus ad hoc.
Complex conditions can often described more easily in a logical form, and this makes integration of with symbolic planner easier.

PRS-Lite: Management process involves determining when to activate/deactivate behaviors as part of execution of a task, as well as coordinating them with other activities in the system. It provides the smooth integration of goal-driven and event-driven activity, while remaining responsive to unexpected changes in the world. The representational basis of PRS-Lite is the activity schema, a parameterized finite-state-machine whose arcs are labelled with goals to be achieved. Each schema embodies procedural knowledge of how to attain some objective via sequence of subgoals, perceptual checks, primitive actions, and behaviors.

When schemas instantiated -> they will be intentions. The details are inside, they can invoke other instentians, block behaviors etc. For example, there is a plan-and-execute schema that first invokes a topological path planner to generate a plan, then launches both an intention to execute the plan and a monitoring intention that oversees the execution to determine when problems have arisen.

Coherence: Reactive behaviors take their inputs directly from sensor readings and more goal-directed behaviors such as wall-following can often benefit from using artifacts (corridor artifact). This is especially true when sensors give only sporadic and uncertain information.

Feature extraction and anchoring is described in detail. Anchoring is a process where current artifacts are kept coherent with the environment by matching against features or object hypothesis. For example, when artifact not found but some features are matched (like corner of a doorway), it may be assumed that artifact is found. For example, artifact say something about the orientation of the corridor is false in LPS, it can be adjusted by sonar readings.
Anchoring thus helps to correct uncertain prior information by using perception, and keep the robot localized on a map as it navigates. However, when environment is complex, features are complex, and robot may be confused.

TOP