Russell and Norvig: an active, artificial agent. continuum of physical configurations and motions

Transcription

1 Chapter 8 Robotics Christian Jacob jacob@cpsc.ucalgary.ca Department of Computer Science University of Calgary 8.5 Robot Institute of America defines a robot as a reprogrammable, multifunction manipulator designed to move material, parts, tools or specific devices through variable programmed motions for the performance of a variety of tasks. Russell and Norvig: an active, artificial agent whose environment is the physical world. Robots differ from Softbots whose environment consists of computer systems, databases and networks. The Physical World The physical world is very demanding, it is: inaccessible - sensors are imperfect, only stimuli that are near the agent can be perceived. nondeterministic - a robot needs to deal with uncertainty nonepisodic - effects of an action change over time dynamic - robot needs to decide when to think and when to act immediately continuous - states and actions are drawn from a continuum of physical configurations and motions 8.5 Manufacturing and materials handling

2 Gofer robots Gofer robots Carnegie Mellon s Nomad Bell & Howell Mailmobile Hazardous environments Hazardous environments Dante II Frame Walking Robot Lunokhod Moon Robot Telepresence and virtual reality Telepresence and virtual reality The Wheelbarrow, a bomb disposal robot Advanced Tethered Vehicle (ATV)

3 Telepresence and virtual reality Augmentation of human abilities Advanced Robot and Telemanipulator System for Minimal Invasive Surgery (ARTEMIS) Sigourney Weaver in the movie Aliens Augmentation of human abilities 8.5 General Electric s Walking Truck Effectors: Tools for Action Locomotion Manipulation Sensors: Tools for perception Proprioception Force Sensing Tactile Sensing Sonar Camera Data Effectors: Locomotion Carnegie Mellon s Ambler

4 Effectors: Locomotion Effectors: Manipulation Degrees of Freedom MIT s 3D Hopper Sensors: Proprioception Haptics MIT Touch Lab MIT s Spring Flamingo Sensors: Force Sensing Sensors: Tactile Sensing MIT s Phantom MIT s Planar Grasper

5 Sensors: Sonar Sensors: Light Sensors Grey Walter s Tortoise: Machina Speculatrix ActivMedia s Peoplebot Sensors: Camera Data Sensors: Camera Data MIT s Fast Eye Gimbals The Beast: Johns Hopkins University 8.5 Architectures The architecture of a robot defines how the job of generating actions from percepts is organized. It is basically the control mechanism of the robot. Classical Architecture

6 Architectures: Classical Architecture A robot with classical architecture is given a number of low-level actions (LLAs). It then uses these LLAs to reason about the effects of performing a sequence of these LLAs. The problem with this is that due to things like wheel slippage and measurement errors any lengthy sequence of actions is prone to fail. Architectures Classical Architecture SRI s Shakey Architectures: The process of deliberating is often too expensive to generate real-time behavior. Situated automata do not explicitly reason, they operate by reflex. A situated automata has two parts: - The first collects sensor inputs and updates the state register accordingly. - The second looks at the state register and calculates output (actions). Thus a situated automata does not plan, it just does whatever it knows to do given the state it is in. Architectures SRI s Flakey 8.5 Configuration Spaces Configuration Space is the path where a robot can move from one position to another. Generalized configuration space Recognizable sets

7 Configuration Spaces Generalized configuration space Generalized configuration space includes other objects as part of the configuration, which could be movable, variable in shapes (i.e. scissors or staples), or deformable (i.e., string or paper). Configuration Spaces Recognizable Sets Includes envelope of possible configurations 8.5 Cell decomposition Skeletonization Fine-motion (bounder-error) planning Landmark-based navigation Online algorithms Cell decomposition Breaks continuous space into a finite number of discrete search problems Skeletonization methods Computes a one-directional skeleton (subset) of the configuration space, yielding an equivalent graph search problem Bell & Howell Mailmobile

8 Fine-motion (Bounded-error) Planning This methods assume bounds on sensor and actuator uncertainty, and in some cases can compute plans that are guaranteed to succeed even in the face of severe actuator errors partial knowledge of the environment is known to the system most of the planning is done offline used for planning small, precise motions of assembly robots Online algorithm The robot makes decisions at run time (no need for offline planning This method assumes that the environment is completely unknown The robot cannot see anything. It can only sense a boundary LOGO Robot Mars Exploration Rovers The robot is equipped with a position sensor and knows the location of its goal. Mars Pathfinder Sojourner Robo Sapiens? Seymour Papert [Kurzweil, 1990] 8. Landmark-based navigation This method assumes that some regions exist in which the robot location can be pinpointed using landmarks, whereas outside those regions it may have only orientation information. This method is both sound and complete The plan has at most n steps if there are n landmarks Configuration Spaces Robo sapiens? What are Robots Good For? What are Robots Made Of? Architectures Robotics Robo Sapiens? KISMET MIT [Menzel and D Aluisio, 2000]

9 Robo Sapiens? WABOT, theorgan Player Ichiro Kato, Waseda-University, Tokyo [Kurzweil, 1990] References P. Menzel and F. D Aluisio (2000). Robo sapiens Evolution of a New Species.. Cambridge, MA, MIT Press. Kurzweil,, R. (1990). The Age of Intelligent Machines. Cambridge, MA, MIT Press.