Transcript Robotics - People.cs.uchicago.edu

Robotics
CSPP 56553
Artificial Intelligence
March 10, 2004
Roadmap
• Robotics is AI-complete
– Integration of many AI techniques
• Classic AI
– Search in configuration space
• (Ultra) Modern AI
– Subsumption architecture
• Multi-level control
• Conclusion
Mobile Robots
Robotics is AI-complete
• Robotics integrates many AI tasks
– Perception
• Vision, sound, haptics
– Reasoning
• Search, route planning, action planning
– Learning
• Recognition of objects/locations
• Exploration
Sensors and Effectors
• Robotics interact with real world
• Need direct sensing for
– Distance to objects – range finding/sonar/GPS
– Recognize objects – vision
– Self-sensing – proprioception: pose/position
• Need effectors to
– Move self in world: locomotion: wheels, legs
– Move other things in world: manipulators
• Joints, arms: Complex many degrees of freedom
Real World Complexity
• Real world is hardest environment
– Partially observable, multiagent, stochastic
• Problems:
– Localization and mapping
• Where things are
• What routes are possible
• Where robot is
– Sensors may be noisy; Effectors are imperfect
– Don’t necessarily go where intend
– Solved in probabilistic framework
Navigation
Application: Configuration Space
• Problem: Robot navigation
– Move robot between two objects without
changing orientation
– Possible?
• Complex search space: boundary tests, etc
Configuration Space
• Basic problem: infinite states! Convert to finite
state space.
• Cell decomposition:
– divide up space into simple cells, each of which can be
traversed “easily" (e.g., convex)
• Skeletonization:
– Identify finite number of easily connected points/lines
that form a graph such that any two points are
connected by a path on the graph
Skeletonization Example
• First step: Problem transformation
– Model robot as point
– Model obstacles by combining their perimeter
+ path of robot around it
– “Configuration Space”: simpler search
Navigation
Navigation
Navigation as Simple Search
• Replace funny robot shape in field of funny
shaped obstacles with
– Point robot in field of configuration shapes
• All movement is:
– Start to vertex, vertex to vertex, or vertex to goal
• Search: Start, vertices, goal, & connections
• A* search yields efficient least cost path
Online Search
• Offline search:
– Think a lot, then act once
• Online search:
–
–
–
–
Think a little, act, look, think,..
Necessary for exploration, (semi)dynamic env
Components: Actions, step-cost, goal test
Compare cost to optimal if env known
• Competitive ratio (possibly infinite)
Online Search Agents
• Exploration:
– Perform action in state -> record result
– Search locally
• Why? DFS? BFS?
• Backtracking requires reversibility
– Strategy: Hill-climb
• Use memory: if stuck, try apparent best neighbor
• Unexplored state: assume closest
– Encourages exploration
Acting without Modeling
• Goal: Move through terrain
• Problem I: Don’t know what terrain is like
– No model!
– E.g. rover on Mars
• Problem II: Motion planning is complex
– Too hard to model
• Solution: Reactive control
Reactive Control Example
• Hexapod robot in rough terrain
• Sensors inadequate for full path planning
• 2 DOF*6 legs: kinematics, plan intractable
Model-free Direct Control
• No environmental model
• Control law:
– Each leg cycles: on ground; in air
– Coordinate so that 3 legs on ground (opposing)
• Retain balance
• Simple, works on flat terrain
Handling Rugged Terrain
• Problem: Obstacles
– Block leg’s forward motion
• Solution: Add control rule
– If blocked, lift higher and repeat
– Implementable as FSM
• Reflex agent with state
FSM Reflex Controller
Retract, lift
higher
yes
no
S3
Move
Forward
Stuck?
S4
Set
Down
Lift up
S2
Push back
S1
Emergent Behavior
• Reactive controller walks robustly
– Model-free; no search/planning
– Depends on feedback from the environment
• Behavior emerges from interaction
– Simple software + complex environment
• Controller can be learned
– Reinforcement learning
Subsumption Architecture
• Assembles reactive controllers from FSMs
– Test and condition on sensor variables
– Arcs tagged with messages; sent when traversed
• Messages go to effectors or other FSMs
– Clocks control time to traverse arc- AFSM
– E.g. previous example
• Reacts to contingencies between robot and env
• Synchronize, merge outputs from AFSMs
Subsumption Architecture
• Composing controllers from composition of
AFSM
– Bottom up design
• Single to multiple legs, to obstacle avoidance
– Avoids complexity and brittleness
• No need to model drift, sensor error, effector error
• No need to model full motion
Subsumption Problems
• Relies on raw sensor data
– Sensitive to failure, limited integration
– Typically restricted to local tasks
• Hard to change task
– Emergent behavior – not specified plan
• Hard to understand
– Interactions of multiple AFSMs complex
Solution
• Hybrid approach
– Integrates classic and modern AI
• 3 layer architecture
– Base reactive layer: low-level control
• Fast sensor action loop
– Executive (glue) layer
• Sequence actions for reactive layer
– Deliberate layer
• Generates global solutions to complex tasks with planning
• Model based: pre-coded and/or learned
• Slower
• Some variant appears in most modern robots
Conclusion
• Robotics as AI microcosm
– Back to PEAS model
• Performance measure, environment, actuators, sensors
– Robots as agents act in full complex real world
• Tasks, rely on actuators and sensing of environment
– Exploits perceptions, learning, and reasoning
– Integrates classic AI search, representation with
modern learning, robustness, real-world focus