Intelligent Robotics and Perception

Intelligent Robotics and Perception
Course Introduction and Overview Instructor: Tucker Balch Wonderful to be here! This is work led by me in collaboration with my graduate students at CMU I am going to focus on work begun over the last year that I plan to continue in the future General -- collaborations

Course Objectives Know what it takes to make a robust autonomous robot work: Sense/Think/Act Understand the important, approaches, research issues and challenges in autonomous robotics. Know how to program an autonomous robot. Just showed you a robotic example of agents interacting with agents we want to develop CS to observe and model multiagent systems we need an example system Collaboration Ant algorithms network routing robot navigation scheduling Interested in application to CS and biology Veloso modeling soccer agents for a while Observing ants just began recently Biologists currently use pencil and paper Tucker Balch BORG Lab CS 4630

Example Videos Tucker Balch BORG Lab CS 4630

How we’re going to do it Read Program: Talk & think in class
Text “Introduction to AI Robotics.” Supplementary papers. Program: Simulated robots Real robots Talk & think in class Our hypothesis is that these algorithms will apply to the study of any multiagent system. Work in progress Give you a picture of our intermediate results and where we want to go Multidisciplinary Multi-disciplinary vision agent Tucker Balch BORG Lab CS 4630

Your Responsibility www.cc.gatech.edu/~tucker/courses/cs4630
Read and understand class policies list: Check your mail several times a week Tucker Balch BORG Lab CS 4630

Evaluation 3 Exams and Final 50% 3 Projects and Final Project 50%
Grading A B C D Other F Brief, short Tucker Balch BORG Lab CS 4630

Final Project Tucker Balch BORG Lab CS 4630

CS 4630: Intelligent Robotics and Perception
History of Intelligent Robotics (Chapter 1) Instructor: Tucker Balch Wonderful to be here! This is work led by me in collaboration with my graduate students at CMU I am going to focus on work begun over the last year that I plan to continue in the future General -- collaborations

Course Objectives Know what it takes to make a robust autonomous robot work: Sense/Think/Act Understand the important, approaches, research issues and challenges in autonomous robotics. Know how to program an autonomous robot. Just showed you a robotic example of agents interacting with agents we want to develop CS to observe and model multiagent systems we need an example system Collaboration Ant algorithms network routing robot navigation scheduling Interested in application to CS and biology Veloso modeling soccer agents for a while Observing ants just began recently Biologists currently use pencil and paper Tucker Balch BORG Lab CS 4630

Example Videos Tucker Balch BORG Lab CS 4630

How we’re going to do it Read Program: Talk & think in class
Text “Introduction to AI Robotics.” Supplementary papers. Program: Simulated robots Real robots Talk & think in class Our hypothesis is that these algorithms will apply to the study of any multiagent system. Work in progress Give you a picture of our intermediate results and where we want to go Multidisciplinary Multi-disciplinary vision agent Tucker Balch BORG Lab CS 4630

Your Responsibility www.cc.gatech.edu/~tucker/courses/cs4630
Read and understand class policies NOT! list: Check your mail several times a week Tucker Balch BORG Lab CS 4630

Evaluation 3 Exams and Final 50% 3 Projects and Final Project 50%
Grading A B C D Other F Brief, short Tucker Balch BORG Lab CS 4630

Final Project Tucker Balch BORG Lab CS 4630

“Intelligent” Robotics
Sense/Think/Act “AI” view “get the computer (robot) to do things that, for now, people are better at” Symbol systems’ hypothesis – intelligence is concerned with the machinery of manipulating symbols “Reactive” view “elephants don’t play chess” Chess is easy – moving around is hard Tucker Balch BORG Lab CS 4630

What Can Robots Be Used For?
Manufacturing 3 Ds Dirty Dull Dangerous Space Satellites, probes, planetary landers, rovers Military Agriculture Construction Entertainment Consumer? Tucker Balch BORG Lab CS 4630

History of Intelligent Robotics
First remote manipulators for hazardous substances 1950s Industrial manipulators: “reprogrammable and multi-functional mechanism designed to move materials, parts, tools…” Closed loop control Tucker Balch BORG Lab CS 4630

History Continued 1955 – term “AI” coined 1960s manufacturing robots
Automatic guided vehicles (AGVs) Precision, repeatability Emphasis on mechanical aspects 1970s Planetary landers Machine vision research expands 1980s Black factory First intelligent autonomous robots: Shakey, Stanford Cart, etc Tucker Balch BORG Lab CS 4630

History Continued 1990s 2000s Symbolic AI/Robotics stalls
Reactive/Behavior-based robotics emerges 2000s ? Tucker Balch BORG Lab CS 4630

Teleoperation Human controls robot remotely Considerations
Hazardous materials Search and rescue Some planetary rovers Considerations Feedback (video, tactile, smell?) User interfaces (cognitive fatigue, nausea) Time/distance Tucker Balch BORG Lab CS 4630

Telepresence Remote embodiment (VR) Considerations
Greater sensor feedback High bandwidth Tucker Balch BORG Lab CS 4630

Semi-autonomous Control
“Supervisory” control Fusion of human commands and autonomous control Delegate some aspects to computer Easier to do in the short term Can be “trusted” Predator (first robot to fire a weapon in combat) Tucker Balch BORG Lab CS 4630

Full Autonomous Control
Tucker Balch BORG Lab CS 4630

Assignments Read Chapter 1 (Weds) Read Paper on Web (Fri) Tucker Balch
BORG Lab CS 4630

Planning (Chapter 2) Instructor: Tucker Balch Wonderful to be here! This is work led by me in collaboration with my graduate students at CMU I am going to focus on work begun over the last year that I plan to continue in the future General -- collaborations

The Planning View Sense/Think/Act
In the planning view “thinking” means to build a model of the world, and deliberate over the model before acting. Tucker Balch BORG Lab CS 4630

General Approach to Planning
Define Possible states (e.g. situations) Operators (actions) that move the robot from one state to another Operator costs Problem Find some sequence of operators that move robot from start state to goal state Optimize? Our hypothesis is that these algorithms will apply to the study of any multiagent system. Work in progress Give you a picture of our intermediate results and where we want to go Multidisciplinary Multi-disciplinary vision agent Tucker Balch BORG Lab CS 4630

Example Problem: Navigation
Goal Start Tucker Balch BORG Lab CS 4630

Navigation State: location (x,y) Operators: move N, S, E, W
Costs: 1 per move Start Goal Goal Start Tucker Balch BORG Lab CS 4630

Contest! Rules Objective: reach goal first
Cannot “sense” obstacles until next to them May “teleport” to any location you have been before “move” Must devise algorithm first, then stick to it Tucker Balch BORG Lab CS 4630

Assignments Read Chapter 1 (Weds) Read Paper on Web (Fri)
Read Chapter 2 (Fri) Tucker Balch BORG Lab CS 4630

The Planning View Sense/Think/Act
In the planning view “thinking” means to build a model of the world, and deliberate over the model before acting. Tucker Balch BORG Lab CS 4630

General Approach to Planning
Define Possible states (e.g. situations) Operators (actions) that move the robot from one state to another Operator costs Problem Find some sequence of operators that move robot from start state to goal state Optimize? Our hypothesis is that these algorithms will apply to the study of any multiagent system. Work in progress Give you a picture of our intermediate results and where we want to go Multidisciplinary Multi-disciplinary vision agent Tucker Balch BORG Lab CS 4630

Strips Used to control Shakey Related to GPS “means-ends analysis”
Reduce differences between current and goal states Example task: get from Tampa to SAIL Tucker Balch BORG Lab CS 4630

Differences, Preconditions & Post-conditions
Operator Pre-conditions Add-list Delete-list d<=200 fly(X,Y) At Y At airport At X 100<d<200 ride_train(X,Y) At station d<=100 drive_rental(X,Y) drive_personal(X,Y) At home d<1 walk(X,Y) Tucker Balch BORG Lab CS 4630

Strips Given World model representation
Difference table, operators, preconditions, postconditions Difference evaluator Tucker Balch BORG Lab CS 4630

Strips Compute diff between goal and initial state. If no difference, terminate If there is a difference reduce the difference by selecting the first operator whose add-list negates the difference Examine the preconditions to see if a set of bindings can be obtained that are all true. If not, make the first false precondition, make it the new goal. Store original goal on the stack. When all preconditions for an operator match, push the operator on the plan stack and update a copy of the world model. Return to the operator with the failed precondition so it can apply its operator or recurse on another failed precondition. Tucker Balch BORG Lab CS 4630

Assignments Finish Chapter 2 Start Chapter 3 (complete by Friday)

The Hierarchical Paradigm
World Model Planner Sensors Actuators The Environment Tucker Balch BORG Lab CS 4630

Hierarchical Architectures
What is a robot architecture? A method of implementing a paradigm, of embodying the principles in some concrete way. Example hierarchical architectures Nested Hierarchical Controller RCS Tucker Balch BORG Lab CS 4630

Nested Hierarchical Controller (Meystel)
Mission Planner Goal World Model Navigator Path Pilot Turn/Drive Low Level Controller Sensors The Environment Tucker Balch BORG Lab CS 4630

NHS Interleaves planning and acting (unlike Strips)
Revises plan if new information comes available Disadvantage Only appropriate for navigation tasks Decomposition is specific to navigation Tucker Balch BORG Lab CS 4630

Real-time Control System (Albus)
Sense Model/Plan Act Time scale The Environment Tucker Balch BORG Lab CS 4630

Difficulties with Planning/Hierarchical Architectures
Closed world assumption: The world model contains everything the robot needs to know Even if closed world assumption is true Models are huge Hard to maintain Poor or no handling of uncertainty and failure Tucker Balch BORG Lab CS 4630

The Frame Problem How to maintain the world model? Tucker Balch
BORG Lab CS 4630

Assignments Chapter 3 by Friday Tucker Balch BORG Lab CS 4630

Biological Foundations (Chapter 3) Instructor: Tucker Balch Wonderful to be here! This is work led by me in collaboration with my graduate students at CMU I am going to focus on work begun over the last year that I plan to continue in the future General -- collaborations

1970s & 80s. Progress in Robotics Slow
Moravec: Stanford Cart Stereo vision Arbib Investigated models of animal intelligence Arkin & Brooks Used models of animal intelligence to control robots. Tucker Balch BORG Lab CS 4630

Why Study Biology for Robotics?
Existence proof of successful systems Animals live in the “open world” Simple animals exhibit intelligence with small brains Thought question: What if a successful robot is developed on the basis of a biological theory that turns out to be wrong? Tucker Balch BORG Lab CS 4630

Animal Behaviors Fundamental building block of natural intelligence.
Behavior: “mapping of sensory inputs to a pattern of motor actions.” Several types of behaviors: Reflexive & stimulus/response (hardwired) Reactive (learned, then used without thought) Conscious behavior (deliberative) Tucker Balch BORG Lab CS 4630

Reflexive Behaviors Reflexes Taxes Fixed-action behaviors
Output lasts only as long as the stimulus remains. Response is proportional to stimulus Taxes Response is to move to a particular orientation (e.g. towards light) Fixed-action behaviors Response continues for a longer duration than the stimulus (e.g. fleeing) Tucker Balch BORG Lab CS 4630

Four Ways to Acquire Behavior
Lorenz and Tinbergen Innate: born with it. E.g. arctic tern fledglings peck at red blobs when hungry. Sequence of innate behaviors: E.g. digger wasp: mate, build nest, lay eggs. Order is triggered in steps Innate sequence with memory Learn Some skills are “passed down” Tucker Balch BORG Lab CS 4630

Innate Releasing Mechanisms
Lorenz and Tinbergen Specific stimulus initiates or triggers the pattern of action E.g. predator present -> flee Supports implicit chaining Tucker Balch BORG Lab CS 4630

Assignments Take home quiz due Weds Project due Friday Tucker Balch
BORG Lab CS 4630

Biological Foundations (Chapter 3) Instructor: Tucker Balch Wonderful to be here! This is work led by me in collaboration with my graduate students at CMU I am going to focus on work begun over the last year that I plan to continue in the future General -- collaborations

Animal Behaviors Fundamental building block of natural intelligence.
Behavior: “mapping of sensory inputs to a pattern of motor actions.” Several types of behaviors: Reflexive & stimulus/response (hardwired) Reactive (learned, then used without thought) Conscious behavior (deliberative) Tucker Balch BORG Lab CS 4630

Reflexive Behaviors Reflexes Taxes Fixed-action behaviors
Output lasts only as long as the stimulus remains. Response is proportional to stimulus Taxes Response is to move to a particular orientation (e.g. towards light) Fixed-action behaviors Response continues for a longer duration than the stimulus (e.g. fleeing) Tucker Balch BORG Lab CS 4630

Innate Releasing Mechanisms
Lorenz and Tinbergen Specific stimulus initiates or triggers the pattern of action E.g. predator present -> flee Supports implicit chaining Tucker Balch BORG Lab CS 4630

Concurrent Behaviors Must somehow be arbitrated Equilibrium
Example: squirrel, flee or eat? Dominance (Winner take all) Example: hungry or sleepy. Can do one but not both. Cancellation One cancels the other – sickleback fish Tucker Balch BORG Lab CS 4630

Perception Two functions Release behaviors Accomplish behaviors

Ecological Approach to Perception
J.J. Gibson Ecological (situated) perception. Perception co-evolves with environment (example: bees). Perception can be fooled (example: fishing lures). “The world is its own best representation.” (quote is actually due to Brooks). Affordances – perceivable potentialities of the environment for action. (example: what is a chair?) Tucker Balch BORG Lab CS 4630

Assignments Take home quiz due Today Project due Friday Tucker Balch
BORG Lab CS 4630

Tucker Balch Wonderful to be here! This is work led by me in collaboration with my graduate students at CMU I am going to focus on work begun over the last year that I plan to continue in the future General -- collaborations

Foraging Robots (1994) Balch, et al, AI Magazine, 1995. Tucker Balch
Georgia Institute of Technology AMiRE October, 2001

Foraging Robots (1997) Balch, AI Magazine, 1997.
Balch, Autonomous Robots, 2000. Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Foraging Robots (2001) Stroupe & Balch, submitted.
Emery & Balch, submitted. Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Foraging Robots ( ) Basic Foraging Heterogeneous Communication & Cooperation Behaviors & Coordination Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Behavior-Based Control: Biological Inspiration
Food Barrier X Frog Arbib, 198? Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Tucker Balch Georgia Institute of Technology AMiRE October, 2001

X Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Motor Schemas Multiple independent processes each generate a vector combined by weighted summation Computationally simple and fast Enables design by composition. (Arkin 1989) Related to artificial potential fields Khatib (85), Krogh (84), Payton (89), Singh (98) Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Motor Schemas: Move to Goal

Motor Schemas: Avoid Obstacle

Motor Schemas: Avoid Obstacle + Move to Goal

Behavior Based Pushing

CS 4630: Intelligent Robotics and Perception Case Study: Motor Schema-based Design Chapter 5
Tucker Balch

What We’ve Covered History of Intelligent Robotics (Chapter 1)
Hierarchical paradigm (Chapter 2) Biological basis for behavior-based control (Chapter 3) Overview of behavior based control (Chapter 4) Subsumption architecture (Chapter 4) Motor schema-based control (Chapter 4) Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Upcoming Today: case study of behavior-based control for multirobot team. Friday: TeamBots tutorial, new project assignment Monday: Midterm Exam Weds: Begin Chapter 5 (Sensing) Friday: Guest Lecture (Koenig) Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Social Potentials Balch & Arkin, IEEE Transactions on Robotics and Automation, 2000 Tucker Balch Georgia Institute of Technology AMiRE October, 2001

The Multi-Foraging Task

Foraging Robots (1997) Balch, AI Magazine, 1997.
Balch, Autonomous Robots, 2000. Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Behavioral Sequencing
at red bin Aquire Red see red Deliver Red have red ~have red Search ~see red Acquire Blue Deliver Blue have blue see blue at blue bin ~see blue ~have blue Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Performance as Team Size Increases

Problem: Inter-Robot Interference

Heterogeneous Strategy 1: Specialization

Heterogeneous Strategy 2: Territorial

Performance Comparison

Are Diversity and Performance Correlated?
Need a measure of robot team diversity Approach: information theory Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Diversity and Performance Negatively Correlated in Foraging

Diversity and Performance Positively Correlated in Soccer
Homogeneous Team Heterogeneous Team Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Where We are Learning Algorithms Behavioral Sequence Representation
Real-time Video Processing Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Observing and Modeling Live Multi-Agent Systems
Motivation Our agents should act intelligently in the presence of other agents: humans, external agents, adversaries Social insects: Rich, multiagent interactions Adversarial/territorial behaviors Real biology in collaboration with entomologists Just showed you a robotic example of agents interacting with agents we want to develop CS to observe and model multiagent systems we need an example system Collaboration Ant algorithms network routing robot navigation scheduling Interested in application to CS and biology Veloso modeling soccer agents for a while Observing ants just began recently Biologists currently use pencil and paper Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Research Goal: Develop Algorithms That Enable
Simultaneous tracking of all the individuals in a colony Recognition of individual and colony behaviors Learning of new single and multi-agent behavior models Application of the models to multi-agent software and robotic systems Our hypothesis is that these algorithms will apply to the study of any multiagent system. Work in progress Give you a picture of our intermediate results and where we want to go Multidisciplinary Multi-disciplinary vision agent Tucker Balch Georgia Institute of Technology AMiRE October, 2001

The complexity of ant society
Holldobler & Wilson, 1990 Gordon, 1999 Brief, short Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Video of ant behaviors Tucker Balch Georgia Institute of Technology
AMiRE October, 2001

Finding Ants In Images (1)
CMVision: Color-based tracking Initially developed for tracking soccer robots Classifies and segments regions according to color 100s of regions, 32 colors, 30Hz, low cost 2 integer ANDs versus 196 comparisons Bruce, Balch & Veloso, IROS-2000 Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Finding ants in images (2)
Approach: background subtraction Enables classification by color and motion Bij = (1 - )Bij + Iij Calculate background using established approach What is unique about our work, we only do the subtraction where there are ants 32 colors separate alphas for different colors Background Image Current Image Movement Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Associating observations with individuals
The association problem Best optimal algorithm O(n3) Greedy approach O(n2) Noisy data presents additional challenges Splitting, merging, drop-outs, pop-ups Current approach “Greedy agents” leverage domain knowledge Future Parallel implementations, Bayesian techniques (e.g. Xiang & Lesser), radar tracking techniques N == number of ants Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Analyzing the Spatial Behavior of a Colony

No Food Available Tucker Balch Georgia Institute of Technology AMiRE
October, 2001

Food Available Tucker Balch Georgia Institute of Technology AMiRE
October, 2001

Vector Representation
Work with Rande Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Recognition Task: Right Turn, Left Turn, Straight?
Approach: Average turning angles over a window Classify turns according to average: if A < -, right turn if A > , left turn otherwise, straight n is the window size Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Example Tucker Balch Georgia Institute of Technology AMiRE
October, 2001

Recognizing Behavior from Movement Traces
Hypothesis: Observed movement features considered over time can be used to classify the behavior of a physical agent Previous success in observation of soccer agents Hidden Markov Models (Han & Veloso, 1999) Example features binary: towards-food, at-food, towards-home, at-home continuous: velocity, turn-rate, path randomness Example behaviors foraging, patrolling, carrying, recruiting Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Hidden Markov Model Representation
B C S1 S3 S2 0.9 0.1 AAABBBBBBBBBBBBCCABBBBBBBBBCCA Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Hidden Markov Model Representation
B 0.1 C 0.1 A 0.1 B 0.8 C 0.1 A 0.1 B 0.1 C 0.8 S1 S3 S2 0.9 0.1 ACABBBABBBCBBBBCAABBBBABBBBCCA Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Using HMMs for Recognition With the Viterbi Algorithm
AAABBBBBBBBBBBBCCABBBBBBBBBCCA Tucker Balch Georgia Institute of Technology AMiRE October, 2001

inspired by Han & Veloso, 1999

Recognition Algorithms Behavioral Sequence Representation
Learning Algorithms Recognition Algorithms Behavioral Sequence Representation Real-time Video Processing Tucker Balch Georgia Institute of Technology AMiRE October, 2001

Thanks to Zia Khan Manuela Veloso James Bruce Gak Kaminka Pat Riley
Rande Shern Ashley Stroupe DARPA Control of Agent Based Systems (CoABS) Tucker Balch Georgia Institute of Technology AMiRE October, 2001

http://www.cc.gatech.edu/~tucker Theoretical foundations
Build large-scale systems Observe and model multi-agent systems Tucker Balch Georgia Institute of Technology AMiRE October, 2001

www.cc.gatech.edu/~tucker www.cc.gatech.edu/~cprl Tucker Balch
Georgia Institute of Technology AMiRE October, 2001

Observing Ants: Tracking and Analyzing the Behavior of Live Insects
Tucker Balch Collaborative Perception and Robotics Lab Wonderful to be here! This is work led by me in collaboration with my graduate students at CMU I am going to focus on work begun over the last year that I plan to continue in the future General -- collaborations

Intelligent Robotics and Perception

Similar presentations

Presentation on theme: "Intelligent Robotics and Perception"— Presentation transcript:

Similar presentations

About project

Feedback

Log in

Auth with social network:

Intelligent Robotics and Perception

Similar presentations

Presentation on theme: "Intelligent Robotics and Perception"— Presentation transcript:

Similar presentations

About project

Feedback