CURRENT PROJECTS
Distributed Cyber-Physical Architectures for Green Electricity Networks
Sponsor: ARPA-E
Duration: Jan. 2012 - Dec. 2014
Project description: This project proposes a comprehensive, backward compatible, incrementally deployable, and scalable control architecture based on distributed autonomous networked control and the emerging concept of electricity “prosumers”– economically motivated energy ecosystems that can consume, produce or store electricity. The architecture’s distributed intelligence and innovative control design reduces communications requirements and decentralizes control functions. A “flat” electricity industry emerges wherein the actors expose various services related to individual and wide-area objectives, realizing an ultra-reliable internetwork for energy that will enable penetration of high levels of renewable energy and storage, numerous novel value propositions, and energy innovation. This project will advance the development leading to the demonstration of modules integrated in a cohesive framework: an autonomous utility prosumer energy management software, and the industry services computation infrastructure.
Sponsor: ARPA-E
Duration: Jan. 2012 - Dec. 2014
Project description: This project proposes a comprehensive, backward compatible, incrementally deployable, and scalable control architecture based on distributed autonomous networked control and the emerging concept of electricity “prosumers”– economically motivated energy ecosystems that can consume, produce or store electricity. The architecture’s distributed intelligence and innovative control design reduces communications requirements and decentralizes control functions. A “flat” electricity industry emerges wherein the actors expose various services related to individual and wide-area objectives, realizing an ultra-reliable internetwork for energy that will enable penetration of high levels of renewable energy and storage, numerous novel value propositions, and energy innovation. This project will advance the development leading to the demonstration of modules integrated in a cohesive framework: an autonomous utility prosumer energy management software, and the industry services computation infrastructure.
Human-Swarm Interactions for Multi-Robot Teams
Sponsor: Robotics and Intelligent Machines at Georgia Tech
Duration: Dec. 2011 - Aug. 2012
Project description: Imagine that you are surrounded by a million robot mosquitos and you have a single joystick that you can use for interacting with the swarm. How should this interaction be structured? This question is at the heart of this proposal and one can directly identify two naive yet ruefully inappropriate interaction models. The first is to point the joystick at individual mosquitos and “drag” them, one at a time, to some target location/configuration. This would mean that the operator would have to issue one million instructions, which is clearly not the right way of structuring this interaction. The second approach is to select some virtual point, such as the center of mass of the swam, and then move that point. This is easy for the operator. Unfortunately, each robot mosquito needs to know where they are relative to the center of mass, which in turn implies that each mosquito needs to know the location of one million mosquitos, which is clearly not scalable or even desirable. The alternative is that the location of the virtual point is broadcast to all mosquitos, which again is not scalable. What this highly simplified thought experiment reveals is that we need appropriate abstractions of the swarm, e.g., into subsets of agents or with respect to the dynamics, that allows for “effective” interactions. These abstractions must be valid in the sense that they can be controlled, i.e., the operator can use the abstractions to achieve the desired performance, and they must provide the operator with sufficient information about the inner workings of the swarm. Moreover, we need to understand the appropriate communication modalities that humans can use with the swarm (gesture/speech/joystick), and how the swarm can use its decentralized sensors to best perceive input from the human operator.
Sponsor: Robotics and Intelligent Machines at Georgia Tech
Duration: Dec. 2011 - Aug. 2012
Project description: Imagine that you are surrounded by a million robot mosquitos and you have a single joystick that you can use for interacting with the swarm. How should this interaction be structured? This question is at the heart of this proposal and one can directly identify two naive yet ruefully inappropriate interaction models. The first is to point the joystick at individual mosquitos and “drag” them, one at a time, to some target location/configuration. This would mean that the operator would have to issue one million instructions, which is clearly not the right way of structuring this interaction. The second approach is to select some virtual point, such as the center of mass of the swam, and then move that point. This is easy for the operator. Unfortunately, each robot mosquito needs to know where they are relative to the center of mass, which in turn implies that each mosquito needs to know the location of one million mosquitos, which is clearly not scalable or even desirable. The alternative is that the location of the virtual point is broadcast to all mosquitos, which again is not scalable. What this highly simplified thought experiment reveals is that we need appropriate abstractions of the swarm, e.g., into subsets of agents or with respect to the dynamics, that allows for “effective” interactions. These abstractions must be valid in the sense that they can be controlled, i.e., the operator can use the abstractions to achieve the desired performance, and they must provide the operator with sufficient information about the inner workings of the swarm. Moreover, we need to understand the appropriate communication modalities that humans can use with the swarm (gesture/speech/joystick), and how the swarm can use its decentralized sensors to best perceive input from the human operator.
Mid-Level Planning and Control for Articulated Locomoting Systems
Sponsor: DARPA
Duration: Jan. 2011 - Dec. 2012
Project description: Researchers excel at creating demonstrations that work well in controlled laboratory settings. However, the real world has less structure, more uncertainty, and the possibility of dynamically changing conditions. Transitioning demonstrations to this type of environment requires an entirely new set of conceptual and computational techniques and tools. The proposed work will move a class of underactuated mechanical systems, in particular locomoting mechanisms, out of the laboratory and into the real world through new and fundamental contributions to integrated planning and control. The result will be a system that can execute complex motions in cluttered and complicated environments. In fact, the proposed work will constitute fundamental advances in the quest to enable robotic locomotion through environments inhospitable to wheeled vehicles, e.g., confined spaces such as collapsed buildings. These environments inspire other types of mobile mechanisms such as snake robots.
Sponsor: DARPA
Duration: Jan. 2011 - Dec. 2012
Project description: Researchers excel at creating demonstrations that work well in controlled laboratory settings. However, the real world has less structure, more uncertainty, and the possibility of dynamically changing conditions. Transitioning demonstrations to this type of environment requires an entirely new set of conceptual and computational techniques and tools. The proposed work will move a class of underactuated mechanical systems, in particular locomoting mechanisms, out of the laboratory and into the real world through new and fundamental contributions to integrated planning and control. The result will be a system that can execute complex motions in cluttered and complicated environments. In fact, the proposed work will constitute fundamental advances in the quest to enable robotic locomotion through environments inhospitable to wheeled vehicles, e.g., confined spaces such as collapsed buildings. These environments inspire other types of mobile mechanisms such as snake robots.
Motion Grammar Laboratory
Sponsor: National Science Foundation
Duration: Jan. 2011 - Dec. 2013
Project description: Reliable robot coworkers will be required to guarantee the completion of task-level objectives. Factory managers that assign skilled humans to mount a flexible cover on a car door have complete confidence that the task will be achieved. Skilled humans assemble the parts even when they have never before seen the particular cover, are unsure of the exact mounting points and must dig the cover out of a pile of parts. This level of under-specification in the task definition and uncertainty in object positions is beyond the capabilities of existing manipulation planners and control algorithms for robots. Instead of simply grasping and displacing the part from one point to another, humans perform a number of motions such as pushing away obstructing objects, trying initial alignments, re-grasping the part and test-fitting contacts. Not only do these motions appear arbitrary but sometimes humans even give up on one strategy and restart with another. Yet these motions cannot be arbitrary since the space of all displacements is exponentially large relative to the number of environment objects and human poses. Our theory is that human task-level manipulation has significant structure and the proposed Motion Grammar Laboratory will support our collaborative effort uncover it.
Sponsor: National Science Foundation
Duration: Jan. 2011 - Dec. 2013
Project description: Reliable robot coworkers will be required to guarantee the completion of task-level objectives. Factory managers that assign skilled humans to mount a flexible cover on a car door have complete confidence that the task will be achieved. Skilled humans assemble the parts even when they have never before seen the particular cover, are unsure of the exact mounting points and must dig the cover out of a pile of parts. This level of under-specification in the task definition and uncertainty in object positions is beyond the capabilities of existing manipulation planners and control algorithms for robots. Instead of simply grasping and displacing the part from one point to another, humans perform a number of motions such as pushing away obstructing objects, trying initial alignments, re-grasping the part and test-fitting contacts. Not only do these motions appear arbitrary but sometimes humans even give up on one strategy and restart with another. Yet these motions cannot be arbitrary since the space of all displacements is exponentially large relative to the number of environment objects and human poses. Our theory is that human task-level manipulation has significant structure and the proposed Motion Grammar Laboratory will support our collaborative effort uncover it.
MURI: Heterogeneous Unmanned Networked Teams
Sponsor: ONR
Duration: Aug. 2008 - July 2013
Project description: Future Naval Combat Operations and Systems will entail small expeditionary forces with light combat ships, high altitude long endurance vehicles, tactical UAVs, and unmanned underwater vehicles, which must monitor and protect large and complex areas continuously. These Heterogeneous Unmanned Networked Teams (HUNT) must be able to search for potential threats, identify them, track them, and take appropriate action to neutralize them. Because of the dynamic nature of the battleeld, HUNT teams must rapidly allocate and task dierent assets to support time-critical intelligence needs, and re- allocate and retask assets in response to the detection of threats or changes in missions. Pushing the state-of-the-art will require a broader perspective in addressing a variety of hard problems. Sophisticated cooperation mechanisms among intelligent biological organisms, in- cluding humans, will oer critical insight and solution templates for many hard engineering problems. To meet the HUNT challenge we have a assembled an interdisciplinary team of leading researchers who have pioneered work in articial intelligence, vehicle control and robotics, cognitive psychology and human factors, biology, and political economics. Due to the emphasis on heterogeneous cooperation, we have not only assembled a team of vehicle engineers with expertise in traditionally separated domains (UGVs, UAVs, UUVs), but we have also assembled a Biological Think Tank, consisting of experts in principles of coop- eration in traditionally separated biological domains.
Sponsor: ONR
Duration: Aug. 2008 - July 2013
Project description: Future Naval Combat Operations and Systems will entail small expeditionary forces with light combat ships, high altitude long endurance vehicles, tactical UAVs, and unmanned underwater vehicles, which must monitor and protect large and complex areas continuously. These Heterogeneous Unmanned Networked Teams (HUNT) must be able to search for potential threats, identify them, track them, and take appropriate action to neutralize them. Because of the dynamic nature of the battleeld, HUNT teams must rapidly allocate and task dierent assets to support time-critical intelligence needs, and re- allocate and retask assets in response to the detection of threats or changes in missions. Pushing the state-of-the-art will require a broader perspective in addressing a variety of hard problems. Sophisticated cooperation mechanisms among intelligent biological organisms, in- cluding humans, will oer critical insight and solution templates for many hard engineering problems. To meet the HUNT challenge we have a assembled an interdisciplinary team of leading researchers who have pioneered work in articial intelligence, vehicle control and robotics, cognitive psychology and human factors, biology, and political economics. Due to the emphasis on heterogeneous cooperation, we have not only assembled a team of vehicle engineers with expertise in traditionally separated domains (UGVs, UAVs, UUVs), but we have also assembled a Biological Think Tank, consisting of experts in principles of coop- eration in traditionally separated biological domains.
Puppet Choreography and Automated Marionettes
Sponsor: National Science Foundation
Duration: May. 2008 - Apr. 2012
Project description: Puppet choreography is a highly-developed language for controlling mechanically complex marionettes. It has evolved over centuries into a largely standardized form that allows puppeteers to address issues that arise as a result of the complex systems with which they are working. As such, the standardization of the choreography can be thought of as the puppeteers. response to complexity. The proposed work will focus on understanding how puppeteers address complex tasks in their choreographic descriptions of plays and using that understanding to solve questions of importance to computer science and engineering. These goals will be achieved by creating an automated puppet play, which will use insights about puppet choreography to implement embedded control of mechanically complex marionettes engaged in complex coordination tasks. In order to automate a puppet play, there are three key technical hurdles that must be addressed: real-time embedded motion control; strategically handling the complexity associated with coordination of high degree-of-freedom systems; and computer animation and simulation.
Sponsor: National Science Foundation
Duration: May. 2008 - Apr. 2012
Project description: Puppet choreography is a highly-developed language for controlling mechanically complex marionettes. It has evolved over centuries into a largely standardized form that allows puppeteers to address issues that arise as a result of the complex systems with which they are working. As such, the standardization of the choreography can be thought of as the puppeteers. response to complexity. The proposed work will focus on understanding how puppeteers address complex tasks in their choreographic descriptions of plays and using that understanding to solve questions of importance to computer science and engineering. These goals will be achieved by creating an automated puppet play, which will use insights about puppet choreography to implement embedded control of mechanically complex marionettes engaged in complex coordination tasks. In order to automate a puppet play, there are three key technical hurdles that must be addressed: real-time embedded motion control; strategically handling the complexity associated with coordination of high degree-of-freedom systems; and computer animation and simulation.
Pilot Decision Support for Controlling Multiple UAVs
Sponsor: Rockwell Collins, Inc.
Duration: Nov. 2007 - Nov. 2012
Project description: Arguably, the biggest challenge facing the successful deployment of unmanned aerial vehicles (UAVs) in unstructured environments is the level of human involvement needed to carry out the mission. In fact, control and coordination of UAVs typically involve a many-to-one mode of operation in that multiple operators are needed in order to control a single UAV. The explicit purpose of this work is to invert this relationship, i.e. to enable a single pilot to control and coordinate multiple unmanned vehicles. This will allow the pilots to operate much more effectively, and will moreover enable the pilots to coordinate and exploit capability synergies between different UAVs to accomplish the mission objectives more effectively. In particular, the following areas will be covered: 1. Methods for evaluating the progression toward the completion of the mission. 2. Automatic ranking of the vehicles by their impact on the mission objectives, if controlled by the pilot. 3. Autonomous coordination of the vehicles. Moreover, the performance of the algorithms will be demonstrated against operational scenarios in a 3D simulation environment.
FORMER PROJECTS
Sponsor: Rockwell Collins, Inc.
Duration: Nov. 2007 - Nov. 2012
Project description: Arguably, the biggest challenge facing the successful deployment of unmanned aerial vehicles (UAVs) in unstructured environments is the level of human involvement needed to carry out the mission. In fact, control and coordination of UAVs typically involve a many-to-one mode of operation in that multiple operators are needed in order to control a single UAV. The explicit purpose of this work is to invert this relationship, i.e. to enable a single pilot to control and coordinate multiple unmanned vehicles. This will allow the pilots to operate much more effectively, and will moreover enable the pilots to coordinate and exploit capability synergies between different UAVs to accomplish the mission objectives more effectively. In particular, the following areas will be covered: 1. Methods for evaluating the progression toward the completion of the mission. 2. Automatic ranking of the vehicles by their impact on the mission objectives, if controlled by the pilot. 3. Autonomous coordination of the vehicles. Moreover, the performance of the algorithms will be demonstrated against operational scenarios in a 3D simulation environment.
FORMER PROJECTS
SRS: Abstraction-Based Motion Programs for Complex Mechanical Systems
Sponsor: National Science Foundation
Duration: Aug. 2008 - July 2011
Project description: This research project aims at the development of a systematic approach to abstraction-based motion control of complex, physical systems. In particular, it aims at understanding how high-level motion program languages can be made to form a basis for an effective software system for complex, interconnected mechanical systems. For this, novel research will be conducted along the following directions: (1) Motion Description Languages: We will discuss how to construct adequate motion description languages (MDLs) as well as propose a software engine for parsing and compiling such languages based on optimal control techniques. (2) Abstraction-Based Models of Interconnected Systems: A novel, graph-based representation of mechanical systems will be proposed that allows for a compact representation of mechanical systems for simulation and analysis, making it a key component of the proposed software engine. (3) Motion Primitives from Empirical Data: In order to ensure the expressiveness of the MDLs, it is of paramount importance that the motion primitives are sufficiently rich. As such, we propose a method for automatically obtaining such primitives from example data, e.g. generated by human operators.
Sponsor: National Science Foundation
Duration: Aug. 2008 - July 2011
Project description: This research project aims at the development of a systematic approach to abstraction-based motion control of complex, physical systems. In particular, it aims at understanding how high-level motion program languages can be made to form a basis for an effective software system for complex, interconnected mechanical systems. For this, novel research will be conducted along the following directions: (1) Motion Description Languages: We will discuss how to construct adequate motion description languages (MDLs) as well as propose a software engine for parsing and compiling such languages based on optimal control techniques. (2) Abstraction-Based Models of Interconnected Systems: A novel, graph-based representation of mechanical systems will be proposed that allows for a compact representation of mechanical systems for simulation and analysis, making it a key component of the proposed software engine. (3) Motion Primitives from Empirical Data: In order to ensure the expressiveness of the MDLs, it is of paramount importance that the motion primitives are sufficiently rich. As such, we propose a method for automatically obtaining such primitives from example data, e.g. generated by human operators.
ZORRO: The Robotic Fencing System
Sponsor: Robotics and Intelligent Machines at Georgia Tech
Duration: Aug. 2009 - June 2010
Project description: In order to be able to deploy truly safe service and flexible automation robots, robots must be able to interact with humans in effective and meaningful ways. This includes the presence of positive interactions (solving collaborative tasks) and the absence of negative interactions (avoiding human-robot collisions). We propose to examine both of these aspects under a single umbrella, namely human-robot fencing. Although this might sound a bit unusual, it is an application that presents clear challenges for dynamic interactions between robots and humans: (1) Prediction of human intentions and (2) Real-time robot response. These goals will be accomplished with online estimation and hybrid control.
Sponsor: Robotics and Intelligent Machines at Georgia Tech
Duration: Aug. 2009 - June 2010
Project description: In order to be able to deploy truly safe service and flexible automation robots, robots must be able to interact with humans in effective and meaningful ways. This includes the presence of positive interactions (solving collaborative tasks) and the absence of negative interactions (avoiding human-robot collisions). We propose to examine both of these aspects under a single umbrella, namely human-robot fencing. Although this might sound a bit unusual, it is an application that presents clear challenges for dynamic interactions between robots and humans: (1) Prediction of human intentions and (2) Real-time robot response. These goals will be accomplished with online estimation and hybrid control.
Reconfigurable Sensor Networks for Fault-Tolerant In-Situ Sampling
Sponsor: NASA
Duration: Sep. 2006 - Aug. 2009
Project description: The goal of this project is to develop and validate the core technologies needed to enable reconfigurable sensor networks for fault-tolerant in-situ sampling for Earth science applications. The key technologies, which build on prior work done by the proposers, focus on science-driven sensor network diagnosis and topological reconfiguration of sensor networks. Control of reconfigurable sensor networks is fundamentally a difficult problem in which the system must balance issues of power usage, communication versus control, the effectiveness of adapting to the environment as well as to changing science requirements. These issues generally arise due to the limited perception, precision, and range constraints on communication channels that comprise the network. Diagnosis involves identifying and communicating necessary changes in network topology required to achieve science goals and compensate for sensor failure or communication dropouts. Reconfiguration involves physically reconfiguring the network topology based on input from the diagnostic process, in effect establishing a self-adapting sensor network. The novelty of our approach is on the focus of a decentralized versus centralized method of control in which interactions between sensor nodes are modeled topographically and manipulated locally to produce desired global behavior. These technologies will be integrated and demonstrated using a network of mobile sensors applied to a representative Earth science investigation.
Sponsor: NASA
Duration: Sep. 2006 - Aug. 2009
Project description: The goal of this project is to develop and validate the core technologies needed to enable reconfigurable sensor networks for fault-tolerant in-situ sampling for Earth science applications. The key technologies, which build on prior work done by the proposers, focus on science-driven sensor network diagnosis and topological reconfiguration of sensor networks. Control of reconfigurable sensor networks is fundamentally a difficult problem in which the system must balance issues of power usage, communication versus control, the effectiveness of adapting to the environment as well as to changing science requirements. These issues generally arise due to the limited perception, precision, and range constraints on communication channels that comprise the network. Diagnosis involves identifying and communicating necessary changes in network topology required to achieve science goals and compensate for sensor failure or communication dropouts. Reconfiguration involves physically reconfiguring the network topology based on input from the diagnostic process, in effect establishing a self-adapting sensor network. The novelty of our approach is on the focus of a decentralized versus centralized method of control in which interactions between sensor nodes are modeled topographically and manipulated locally to produce desired global behavior. These technologies will be integrated and demonstrated using a network of mobile sensors applied to a representative Earth science investigation.
CSR-EHS: Optimal, Multi-Modal Control of Complex Systems
Sponsor: National Science Foundation
Duration: Aug. 2005 - Aug. 2009
Project description: The technological frontier and performance barrier for control and management of many present-day engineering systems lie in their complexity. This complexity is often due to the high dimensionality and distributed nature of the systems (e.g., manufacturing and transportation systems), the absence of adequate models of the environment in which they operate (autonomous mobile robots), or the vast amounts of data required for their control (airport security systems). Consequently, complexity management has become an essential part of control systems design. An emerging approach to controlling complex systems consists of a decomposition of the control actions into a sequence of modes, each of which is defined for a particular task, operating point, or data source. This results in a hierarchical control structure with an event-driven supervisory control at the higher layer and a time-driven control at the lower layer. Whereas the time-driven feedback laws can be designed by standard control-engineering techniques, the problem of optimally designing the supervisory controllers is by-and-large open due to its inherent high complexity. The central question is how to schedule the various modes in order to optimize the system's performance. Related questions concern the development of real-time algorithms for performance improvement (since optimality may be unrealistic in real time), and the tradeoff between the size of the mode set (complexity) and the system's performance (expressiveness).
Sponsor: National Science Foundation
Duration: Aug. 2005 - Aug. 2009
Project description: The technological frontier and performance barrier for control and management of many present-day engineering systems lie in their complexity. This complexity is often due to the high dimensionality and distributed nature of the systems (e.g., manufacturing and transportation systems), the absence of adequate models of the environment in which they operate (autonomous mobile robots), or the vast amounts of data required for their control (airport security systems). Consequently, complexity management has become an essential part of control systems design. An emerging approach to controlling complex systems consists of a decomposition of the control actions into a sequence of modes, each of which is defined for a particular task, operating point, or data source. This results in a hierarchical control structure with an event-driven supervisory control at the higher layer and a time-driven control at the lower layer. Whereas the time-driven feedback laws can be designed by standard control-engineering techniques, the problem of optimally designing the supervisory controllers is by-and-large open due to its inherent high complexity. The central question is how to schedule the various modes in order to optimize the system's performance. Related questions concern the development of real-time algorithms for performance improvement (since optimality may be unrealistic in real time), and the tradeoff between the size of the mode set (complexity) and the system's performance (expressiveness).
Decentralized Algorithms for Locally Interacting Mobile Robots
Sponsor: US Army Research Office
Duration: Aug. 2005 - Nov. 2008
Project description: The overall objective of this research is to develop communication and control strategies for teams of multiple mobile robots whose knowledge about the environment is constrained. These constraints are generated both from limited perception capabilities, as well as precision and range limitations on the inter-robot communication channels. In particular, it will be investigated how changes in the environment can be dealt with in a decentralized manner through local rules in a predictable manner. This endeavor is driven by a unifying theme in which a number of questions are formulated and solved, ranging from "What global mode of operation is most suited to the current situation?" to "How can this be realized in a systematic manner in the absence of global information?". The effort is mainly focused on the following key areas: 1) Autonomous Formation Selection. 2) Control and Communication Strategies for Local Interactions.
Sponsor: US Army Research Office
Duration: Aug. 2005 - Nov. 2008
Project description: The overall objective of this research is to develop communication and control strategies for teams of multiple mobile robots whose knowledge about the environment is constrained. These constraints are generated both from limited perception capabilities, as well as precision and range limitations on the inter-robot communication channels. In particular, it will be investigated how changes in the environment can be dealt with in a decentralized manner through local rules in a predictable manner. This endeavor is driven by a unifying theme in which a number of questions are formulated and solved, ranging from "What global mode of operation is most suited to the current situation?" to "How can this be realized in a systematic manner in the absence of global information?". The effort is mainly focused on the following key areas: 1) Autonomous Formation Selection. 2) Control and Communication Strategies for Local Interactions.
LAGR: Learning Perception, Controllers and Visual Feature Graphs for Ground Robots
Sponsor: DARPA
Duration: Oct. 2004 - Jan. 2008
Project description: Ground robots should learn from their own experience, and learn from human example. Our research will exploit both opportunities for learning using novel algorithms to learn perception and control. As a baseline framework, we will build a reliable, modular system using well understood perception and control components. Novel learning algorithms will be integrated using a "plug-in" approach to facilitate comparative evaluation. Our main innovative ideas include: 1) Statistical classification and grouping of 3D obstacles and terrain from monocular images. 2) Learning and using visual feature graphs for navigation. 3) Learning controllers from example and experience. 4) Learning statistical models of 3D terrain types and obstacle types.
Sponsor: DARPA
Duration: Oct. 2004 - Jan. 2008
Project description: Ground robots should learn from their own experience, and learn from human example. Our research will exploit both opportunities for learning using novel algorithms to learn perception and control. As a baseline framework, we will build a reliable, modular system using well understood perception and control components. Novel learning algorithms will be integrated using a "plug-in" approach to facilitate comparative evaluation. Our main innovative ideas include: 1) Statistical classification and grouping of 3D obstacles and terrain from monocular images. 2) Learning and using visual feature graphs for navigation. 3) Learning controllers from example and experience. 4) Learning statistical models of 3D terrain types and obstacle types.
CAREER: Linguistic Control of Mobile Robots
Sponsor: National Science Foundation
Duration: Feb. 2003 - Jan. 2008
Project description: When humans instruct each other how to carry out particular tasks, only a limited number of tokenized (linguistic) instructions are used. In contrast to this, classic control theory specifies a control action to be carried out at each time instant. But, in a number of applications, such as semi-autonomous service robots for industrial and domestic use, intelligent appliances, and communication constrained embedded and/or teleoperated devices, the control procedures have a natural, linguistic flavor. The interpretation here is that linguistic control instructions specify particular modes of operation rather than explicit control values. And, the overall objective of this research is to develop and disseminate rational methods for understanding how continuous devices, such as mobile robots, should be controlled using computer generated, linguistic inputs. In particular, it will be studied how these instructions should be defined, selected, and coded in order to minimize the number of bits transmitted from the computer to the robot, while guaranteeing that the system meets its specifications. For this, an information theoretic approach to control theory will be developed, serving as a useful tool not only for source coding of control signals, but also for describing how symbolic instructions should be interpreted and operated on by the continuous systems. Questions concerning what sensors and actuators to use in a given robotic application can be addressed quite elegantly within this framework as well.
Sponsor: National Science Foundation
Duration: Feb. 2003 - Jan. 2008
Project description: When humans instruct each other how to carry out particular tasks, only a limited number of tokenized (linguistic) instructions are used. In contrast to this, classic control theory specifies a control action to be carried out at each time instant. But, in a number of applications, such as semi-autonomous service robots for industrial and domestic use, intelligent appliances, and communication constrained embedded and/or teleoperated devices, the control procedures have a natural, linguistic flavor. The interpretation here is that linguistic control instructions specify particular modes of operation rather than explicit control values. And, the overall objective of this research is to develop and disseminate rational methods for understanding how continuous devices, such as mobile robots, should be controlled using computer generated, linguistic inputs. In particular, it will be studied how these instructions should be defined, selected, and coded in order to minimize the number of bits transmitted from the computer to the robot, while guaranteeing that the system meets its specifications. For this, an information theoretic approach to control theory will be developed, serving as a useful tool not only for source coding of control signals, but also for describing how symbolic instructions should be interpreted and operated on by the continuous systems. Questions concerning what sensors and actuators to use in a given robotic application can be addressed quite elegantly within this framework as well.
Sting Racing: DARPA Urban Grand Challenge
Sponsor: GT CoC, GT ECE, GTRI, SAIC, Telcordia, Duke, Sun Microsystems, Mobile Intelligence
Duration: Jul. 2006 - Nov. 2007
Project description: DARPA's Urban Challenge tests the ability of competing autonomous robots to drive 60 miles in an urban setting in six hours or less. The vehicles must obey the rules of the road, and safely interact with other robot vehicles and other cars driven by people on the course. The competition is scheduled for November 3, 2007. Cash prizes will be awarded to the top three finishers, including $2,000,000 for first place. To win, the driverless robots must have the ability to sense and react to urban traffic environment, including: lane markings, intersections, other vehicles, unexpected road blockages. The robot vehicles must also have the intelligence to select the best route, and decide which vehicle has the right-of-way in normal traffic.
Sponsor: GT CoC, GT ECE, GTRI, SAIC, Telcordia, Duke, Sun Microsystems, Mobile Intelligence
Duration: Jul. 2006 - Nov. 2007
Project description: DARPA's Urban Challenge tests the ability of competing autonomous robots to drive 60 miles in an urban setting in six hours or less. The vehicles must obey the rules of the road, and safely interact with other robot vehicles and other cars driven by people on the course. The competition is scheduled for November 3, 2007. Cash prizes will be awarded to the top three finishers, including $2,000,000 for first place. To win, the driverless robots must have the ability to sense and react to urban traffic environment, including: lane markings, intersections, other vehicles, unexpected road blockages. The robot vehicles must also have the intelligence to select the best route, and decide which vehicle has the right-of-way in normal traffic.
What are the X's Doing?
Sponsor: Robotics and Intelligent Machines at Georgia Tech
Duration: Aug. 2006 - Jun. 2007
Project description: Given empirical examples, obtained for example from biological data or remotely controlled mechanisms, how can mobile robots learn what behaviors to use and when to use them, in order to faithfully reproduce, as well as generalize, the example behaviors? In fact, a basic assumption underlying much of the research in behavior-based robotics is that the modularization of the navigation system into atomic building blocks, or behaviors, is intrinsically sound. In other words, not only does this modularization allow for conceptual complexity reductions, but it is present also in naturally occurring systems. For instance, there is ample evidence that biological navigation systems are modularized as well, in that more complex animal (and human) behaviors can be composed from atomic behaviors. Through this project, we intend to make this assumption explicit by generating atomic control laws, or behaviors, from biological data. Moreover, conditions for when to trigger transitions from one behavior to another will be obtained in an automatic fashion, with the result of producing high-level hybrid automata, containing both continuous dynamics (different behaviors) and discrete logic (transition rules), from the example data.
Sponsor: Robotics and Intelligent Machines at Georgia Tech
Duration: Aug. 2006 - Jun. 2007
Project description: Given empirical examples, obtained for example from biological data or remotely controlled mechanisms, how can mobile robots learn what behaviors to use and when to use them, in order to faithfully reproduce, as well as generalize, the example behaviors? In fact, a basic assumption underlying much of the research in behavior-based robotics is that the modularization of the navigation system into atomic building blocks, or behaviors, is intrinsically sound. In other words, not only does this modularization allow for conceptual complexity reductions, but it is present also in naturally occurring systems. For instance, there is ample evidence that biological navigation systems are modularized as well, in that more complex animal (and human) behaviors can be composed from atomic behaviors. Through this project, we intend to make this assumption explicit by generating atomic control laws, or behaviors, from biological data. Moreover, conditions for when to trigger transitions from one behavior to another will be obtained in an automatic fashion, with the result of producing high-level hybrid automata, containing both continuous dynamics (different behaviors) and discrete logic (transition rules), from the example data.
Communications in Embedded Control Systems
Sponsor: National Science Foundation
Duration: Aug. 2002 - Jul. 2005
Project description: Information sharing in embedded, decentralized control systems is the general area of research that will be pursued in this project. In particular it will be investigated how many bits of information need to be transmitted between different embedded software components in order to make the physical system meet its specifications. For instance, autonomous robotic systems are normally relaying on a variety of heterogeneous sensors, and the question then becomes, which of the sensors do we need?And furthermore, is it possible to compress the data in a systematic way, e.g. using virtual sensors, so that information that is not essential to the current task, or mode of operation, can be discarded? Since the sensory data needed for accomplishing a certain task depends both on the control law, the dynamics of the system, and the complexity of the task, the use of information theoretic tools for embedded, multi-modal control design gives us the means to investigate questions concerning sensor and actuator selection and control mode design in a unified way. The goal of this project is in particular to carry out the following four programs: 1) Model the information theoretic content of the symbolic, computer generated inputs used for controlling continuous, mechanical devises. 2) Select embedded sensors and actuators that make the control system meet its specifications in the presence of bandwidth constraints. 3) Develop coding strategies for compressing the sensory information based on what control application is to be carried out. 4) Apply the theory to the control of multiple autonomous, intelligent robots.
Sponsor: National Science Foundation
Duration: Aug. 2002 - Jul. 2005
Project description: Information sharing in embedded, decentralized control systems is the general area of research that will be pursued in this project. In particular it will be investigated how many bits of information need to be transmitted between different embedded software components in order to make the physical system meet its specifications. For instance, autonomous robotic systems are normally relaying on a variety of heterogeneous sensors, and the question then becomes, which of the sensors do we need?And furthermore, is it possible to compress the data in a systematic way, e.g. using virtual sensors, so that information that is not essential to the current task, or mode of operation, can be discarded? Since the sensory data needed for accomplishing a certain task depends both on the control law, the dynamics of the system, and the complexity of the task, the use of information theoretic tools for embedded, multi-modal control design gives us the means to investigate questions concerning sensor and actuator selection and control mode design in a unified way. The goal of this project is in particular to carry out the following four programs: 1) Model the information theoretic content of the symbolic, computer generated inputs used for controlling continuous, mechanical devises. 2) Select embedded sensors and actuators that make the control system meet its specifications in the presence of bandwidth constraints. 3) Develop coding strategies for compressing the sensory information based on what control application is to be carried out. 4) Apply the theory to the control of multiple autonomous, intelligent robots.
Hybrid, Data-Driven Control in Electronics Manufacturing
Sponsor: Georgia Tech Manufacturing Research Center
Duration: Jun. 2002 - May. 2004
Project description: Closed-loop control of the CBAR SMT-process is an effort that has been sustained over a number of years. This research defines a natural continuation of that work, and the control research aims at the development of robust, data-driven, adaptive, closed-loop control algorithms for the SMT-process. In particular, by closing the loop over the system, the available measurements can be used for adaptively changing the machine parameters in order to achieve good performance on a board-to-board basis. This is an important objective since the dynamics governing the stencil printer is different depending on the direction of the squeegee, which stresses the fact that the machine parameters have to change in between individual boards. Our current research in this area focuses on the generation of provenly convergent iterative algorithms for generating optimal machine parameter sequences in the presence of high noise levels, unknown machine dynamics, and aggressive transient responses. In fact, locally defined closed-loop controllers have been developed, with promising results from both an experimental and a theoretical point of view, that change the squeegee speed and pressure in response to varying solder brick heights.
Sponsor: Georgia Tech Manufacturing Research Center
Duration: Jun. 2002 - May. 2004
Project description: Closed-loop control of the CBAR SMT-process is an effort that has been sustained over a number of years. This research defines a natural continuation of that work, and the control research aims at the development of robust, data-driven, adaptive, closed-loop control algorithms for the SMT-process. In particular, by closing the loop over the system, the available measurements can be used for adaptively changing the machine parameters in order to achieve good performance on a board-to-board basis. This is an important objective since the dynamics governing the stencil printer is different depending on the direction of the squeegee, which stresses the fact that the machine parameters have to change in between individual boards. Our current research in this area focuses on the generation of provenly convergent iterative algorithms for generating optimal machine parameter sequences in the presence of high noise levels, unknown machine dynamics, and aggressive transient responses. In fact, locally defined closed-loop controllers have been developed, with promising results from both an experimental and a theoretical point of view, that change the squeegee speed and pressure in response to varying solder brick heights.