Controller Design

Sends commands to the steering, braking, & throttling actuators. Translates the trajectory path to the goal destination to physical actions.

Motion Model

• A Kinematic model is suitable for smooth turns & slower speeds
• To keep track of the current state based on the previous state & current control inputs
• Autoware already adopted the bicycle kinematic model, as it offers a nice balance of simplicity & accuracy
• Calculates state transitions as a funciton of time
• Derive velocity vector components using trig

• Define physical limitations for:
  • Accelerating (throttle) & decelerating (braking)
  • Steering

Control Scheme

Longitudinal component:

• Controlled variables: throttle & brake inputs.
• Governs the velocity, acceleration, jerk, & higher derivatives
• Use a PID controller for now (typically the primary choice)

Lateral component:

• Controlled variable: steering input
• Governs the steering angle & heading (yaw)

Coupled Control

• If coupled, allow the Longitudinal controller to be dominant over the lateral to prevent a large steering angle at high speed.
• The longitudinal is computed independently, influences lateral, & regulates max steering limit inversely proportional to the vehicle speed.

Model Predictive Control (MPC)

• Best for both normal & rigorous driving conditions
• Can handle multi-input multi-output (can handle both longitudinal & lateral control)
• Constraints can be used for both safety & physical limitations

• Sample Time
  • If too high, response time will be too slow
  • If too small, responses will be over-reactive
  • Suggested sample time: 5% to 10% of rise time
• Prediction Horizon
  • If too small, the vehicle will lack time to react properly
  • If too high, wasteful effort
• Control Horizon
  • The number of time steps that are computed by the optimizer
  • If too short, the optimizer may not return the best possible actions
  • If too long, wasteful effort (only the first couple control actions have a huge impact on the predicted states)
• Weights
  • For prioritizing goals
  • MPC tries to do everything simultaneously, but goals can compete with each other
  • Ex: You can assign higher weight to pose tracking rather than velocity tracking (less important)
• Cost functions can be implemented for both longitudinal & lateral control. Can penalize.
• Optimization chooses the optimal set of control inputs corresponding to the predicted trajectory with lowest cost (while still considering current state, reference state, & constraints). Developer's choice how to implement.

• Ways to reduce computational resources, if necessary:
  • Limit the prediction & control horizons
  • Formulate simpler cost functions
  • Use less detailed motion models
  • Set a high tolerance for optimization convergence

Inputs & Outputs

• Inputs:
  • Trajectory path

• Longitudinal Output:
  • Brake commands
  • Throttle commands
• Lateral Output:
  • Steering angle commands

Note: More inputs & processing nodes will vary on the selected controller & motion model

Links

Control strategies for autonomous vehicles

Parameter adaptive steering control for autonomous vehicles

MPC-based path tracking with PID speed control for autonomous vehicles

Mathworks: Understanding MPC

Mathworks: Three Ways to Speed Up MPC

Mathworks: Developing Longitudinal Controls for a Self-Driving Taxi