From Coursera, State Estimation and Localization for Self-Driving Cars by University of Toronto
https://www.coursera.org/learn/motion-planning-self-driving-cars
Principles of Behaviour Planning
Behaviour Planning
- A behavior planning system plans the set of high level driving actions, or maneuvers to safely achievethe driving mission under various driving situations
- Behavior planner considers:
- Rules of the road
- Static objects around the vehicle
- Dynamic objects around the vehicle
- Planned path must be safe and efficient
Driving maneuvers
- Track Speed: maintain current speed of the road
- Follow leader: match the speed of the leading vehicle and maintain a safe distance
- Decelerate to stop: begin decelerating and stop before a given space
- Stop: remain stopped in the current position
- Merge: join or switch onto a new drive lane
Output of Behavior Planner
- Driving maneuver to be executed
- Set of constraints which must be obeyed by the planned trajectory of the self driving car which include:
- ldeal path
- Speed limit
- Lane boundaries
- Stop locations
- Set of interest vehicles
Input Requirements
- High definition road map
- Mission path
- Localization information
- Perception Information:
- All observed dynamic objects
- Prediction of future movement
- Collision points and time to collision
- All observed static objects
- Road signs
- Occupancy grid
- All observed dynamic objects
Handling an Intersection Scenario Without Dynamic Objects
Scenario Evaluation
- 4 way Intersection
- Two lane
- Stop Sign for every direction
- Be able to travel:
- Through the intersection
- Left at the intersection
- Right at the intersection
- No other dynamic vehicles
Behavior Planning Testing
- Code based tests
- Simulation tests
- Private track tests
- Limited scoped close supervision road tests
Handling an Intersection Scenario with Dynamic Objects
Interaction With Dynamic Objects
- Distance to dynamic object
- distance to the center of any dynamic object
- Distance to collision point
- distance to the collision point with another dynamic object
- Time to collision(TTC)
- time to collision between any two dynamic objects
State Machine States
- Track Speed
- Follow the current speed limit
- Follow Leader
- Match the speed of the dynamic object in front
- Decelerate to Stop
- Stop to a particular point
- Stop
- Stay stopped at the current location
Handling Multiple Scenarios
Hierarchical state machine includes multiple scenarios
Single State Machine
- Single state machine method
- Add transitions
- Add additional transition conditions
- Issues with single state machine method:
- Rule explosion
- Increase in computational time
- Complicated to create and maintain
Hierarchical State Machine-Advantages and Disadvantages
-Advantages:
- Decrease in computational time
- Simpler to create and maintain
- Disadvantages:
- Rule Explosion
- Repetition of many rules in the low level state machines
Advanced Methods for Behaviour Planning
- Fuzzy system
- Reinforcement learning
- Hierarchical Reinforcement Learning
- Model-based Reinforcement Learning
- Machine learning
- Inverse Reinforcement Learning
- End-to-End Approaches