220 likes | 331 Vues
The Alive Project explores the design and control of an autonomous dog agent within a 3D artificial world. This agent, equipped with synthetic vision, aims to satisfy internal motivations and respond to user gestures while pursuing its own agenda. The control architecture consists of multiple levels: motivational, task, and direct, enabling the dog to select actions based on various stimuli. This system incorporates behavior models that allow for optimal action selection and smooth execution of commands, bridging autonomy and directability for enhanced user interaction.
E N D
The Alive ProjectBruce M. Blumberg, MIT Sidney D’Mello Control of Autonomous Agents
Overview: • Agent: Autonomous Dog (Silas T. Dog) • Environment: 3D artificial world and people. • Sensors: Synthetic vision. • Goals: Satisfy internal motivations. Obey the user. • Actions: Respond to gestures of the user. Other basic actions (satisfy drives, etc..)
Goals: • Autonomy: Pursues its own agenda. Senses the environment. Acts on the environment. • Directability: Action selection may be independent of the behavior system. Suggest how the action should be performed.
Multiple Levels of Control • Motivational Level: Change the dog’s current motivation. • Task Level: Supply a high level directive. • Direct Level: Direct control of the motor system.
Behavior Model Design goals: • Optimal action selection model. • Model the effect of external stimuli and internal motivation. • Multiple behaviors may suggest actions and execution preferences. • Maintain a winner takes all architecture. • Support for motivational and task level directions at run time.
Behavior Architecture • Three central parts: • Behavior • Motor Skills • Geometry • Two layers of abstraction: • Controller • Degree’s of Freedom
Geometry • Provides the shapes and transforms. • Issued by the motor system. • Manipulated over time for animation.
Motor System • Components: Controller, Motor Skills, Degree of Freedom • Commands: left, sit, forward, back,…. • Importance: Translates behaviors into commands. Abstraction barrier (forward --- walking, swimming). Provides a generic set of commands (eat, sleep). Provides resource management.
Degrees of Freedom • “Knobs” to modify the underlying geometry. • Used to wag the tail, move a joint,… • Importance: Provides a locking mechanism. Provides an abstraction mechanism.
Motor Skills • Examples: lower head, turning, walking,…. • Can be turned on or off. • Coordinates DOF’s to perform smooth motion. • Spring loaded. • Reduces bookkeeping.
Controller • Abstraction barrier between behavior system and motor skills. • Maps commands to turn on or turn of motor skills. • Example: forward command --- walk, move motor skill. • Command Types: Primary command Secondary command. Meta command.
Hierarchical network of goal-directed entities. Examples: move-to-tree search-for-food knock-on-door Behavior System
Releasing Mechanisms • Filters sensory input. • Identify relevant objects and/or events. • Typically output a continuous value. • May be shared among multiple behavior. • Example: weak stimulus/strong motivation strong stimulus/weak motivation
Internal Variables • Used to model the internal state. • Expressed as continuous values. • Modified by behaviors. • May be shared by multiple behaviors.
Behavior Groups • Groups of mutually inhibiting behaviors. • Loose hierarchicalstructure: • Upper Level: Driven by motivation (engage-in-feeding). • Lower Level: Driven by sensory input (pounce, chew).
Inhibition and Level of Interest • Avalanche Effect. • Insures that only one behavior will be non-zero. • Controls the persistence of behaviors. • Manages the level of interest of behaviors. • Provides a mechanism for the winner takes all arbitration.
Issuing Commands: • Winning behavior and its children: Primary commands. • Losing behaviors: Secondary and Meta commands. • Winning behavior can overrule a secondary command. • Winning behavior can ignore a meta command.
Integration of Directability • Motivational control: Access to internal variables. • Behavior control: Access to Releasing Mechanism. Possible to activate behaviors. • Sensory control: imaginary sensory input. • Motor Control: Shut of the behavior system. Issue secondary or meta commands.
Implementation Specifics • Responds to a dozen user gestures. • 40 Behaviors. • 11 Behavior Groups. • 40 Releasing Mechanisms. • 8 Internal Variables. • 70 Motor Commands.
Conclusions • Action selection algorithm suits needs. • Autonomy and directability are not mutually exclusive. • Blend of autonomy and direction provide more control. • Control architecture can be easily incorporated onto different agents.
References: • Multi-Level Direction of Autonomous Creatures for Real-Time Virtual Environments Bruce M. Blumberg and Tinsley A. Galyean • Expressive Autonomous Cinematography for Interactive Virtual Environments Bill Tomlinson, Bruce Blumberg, Delphine Nain