Robotics for Intelligent Environments
Manfred Huber
[email protected]
March 2002
Robotic Applications in Smart Homes
Control of the physical environment
Automated blinds
Thermostats and heating ducts Automatic room partitioning
Personal service robots
House cleaning
Lawn mowing
Assistance to the elderly and handicapped Office assistants
March 2002 2
What is a Robot ?
Robota (Czech) = A worker of forced labor
Japanese Industrial Robot Association:
A robot is a device with degrees of freedom that can be controlled
Historical Robots include:
Mechanical automata
Motor-driven automata
Computer-controlled robots
March 2002 3
Traditional Robotics
Industrial robot manipulators
Repetitive tasks High speed
Few sensing operations
High precision movements
Pre-planned trajectories and task policies
No interaction with humans
March 2002 4
Robots in Smart Home Environments
Problems of Traditional Robotics:
No sensing Can not handle uncertainty No interaction with humans
Reliance on perfect task information
Complete re-programming for new tasks
March 2002
Current Robots
Design Goals:
Sensor-rich
Flexible
Versatile
Controllable
March 2002
Future Home Robots ?
Peter Menzel / MIT AI Lab
Honda Corp March 2002 7
Challenges for Robots in Intelligent Environments
Control Challenges:
Autonomy in uncertain environments Adaptation and Learning Human-machine interaction
March 2002
Uncertainty in Robot Systems
Sensor Uncertainty:
Sensor readings are imprecise and unreliable
Non-observability:
Various aspects of the environment can not be observed The environment is initially unknown
Action Uncertainty:
Actions can fail Actions have nondeterministic outcomes
March 2002 9
Behavior-Based Robots
Behavior is achieved by combining "reflexes":
Achieves reactivity
Avoids world models Tight coupling of sensors and actions
MIT AI Lab
March 2002
MIT AI Lab http://www-robotics.usc.edu/~maja/robot-video.mpg
10
Probabilistic Robotics
Explicit reasoning about Uncertainty using Bayes filters:
b(st ) p(ot | st ) p(st | st 1, at 1 ) b(st 1) dst 1
Used for:
Localization Mapping Model building
March 2002 11
Hybrid Control Systems
Abstract Planning and Policy Formation Layer
Goal-directed task performance Permits sophisticated reasoning
Reactive Behavior Layer
Ensures basic autonomy
Provides reactivity
Reduces complexity in the planning layer
March 2002 12
Hybrid Control Policies
Finite State Rotation Policy:
March 2002
13
How Many Roboticists does it take to change a Lightbulb ?
March 2002
14
Adaptation and Learning in Robots
Adaptation of Existing Control Policies
Adaptation to changing environments
Adjustment to new user preferences
Learning New Policies
Full autonomy in remote environments Dynamic extension of task repertoire
Learning Sensor Interpretations
Reduction in the amount of data
March 2002 15
Learning Sensory Patterns
Learning to Identify Features
Example Learning Techniques:
Neural Networks Kohonen Maps Unsupervised Clustering
Decision Tree Induction
Chair
March 2002
16
Learning Control Policies
Learning to make Rational Decisions
Challenges:
Learning without supervision Learning in uncertain environments Learning from Human-Machine interaction
March 2002
17
Reinforcement Learning
Learning Control Policies from Reward Signals
Does not require knowledge of the correct policy
Can deal with intermittent, sparse feedback
Q-learning:
Learning an optimal utility function, Q(s, a), for a Markov Decision Processes Q(st-1, a) r + g maxb Q(st, b) Does not require a model
March 2002 18
Learning in Hybrid Control Systems
Policy Acquisition Layer
Learning tasks without supervision
Discrete Event Model Layer
Learning a system model Basic state space compression
Reactive Behavior Layer
Initial competence and reactivity
March 2002 19
Example: Learning to Walk
March 2002
20
Hierarchical Skill Acquisition
Developing Skills Hierarchically
Simplified control policies
Increasingly abstract state spaces Better learning performance
Hierarchical Reinforcement Learning
Learning with abstract actions
Acquisition of abstract task knowledge
March 2002 21
Personal Service Robots
Control Challenges: http://www.cs.cmu.edu/~thrun/movies/pearl-assist.mpg
Robustness requirements Safety and reliability requirements Interaction with humans
Human-Machine interfaces
Application Domains:
Office assistants Home cleanup
March 2002 CMU Robotics Institute http://www/cs/cmu.edu/~thrun/movies/pearl_assist.mpg
Assistance to elderly and handicapped
22
Human-Machine Interfaces and Variable Autonomy
Variable Autonomy
Autonomous operation / learning User operation / teleoperation
Behavioral programming
Following user instructions Imitation
Potential Interfaces:
Keyboard Voice recognition Visual observation
March 2002 23
Human-Machine Interfaces : Teleoperation
Remote Teleoperation
Direct operation of all degrees of freedom by the user
Simple to install Removes user from dangerous areas
Can be exhaustive
Requires insight into the mechanism
Easily leads to operation errors
March 2002 24
"Social" Interactions with Robots
"Attentional" Robots
Focus on the user or task
First step to imitation
"Emotional" Robots
Better acceptance by the user More natural human-machine interaction
MIT AI Lab http://www.ai.mit.edu/projects/cog/Video/kismet/kismet_face_30fps.mpg
Users are more forgiving
March 2002 25
Summary
Robots in Intelligent Environments require:
Autonomous Control
Adaptation and Learning Capabilities
Flexible Human-Machine Interfaces Versatile Mechanisms
March 2002 26
