|
Autoware.Auto
|
|
Model predictive control (MPC) involves solving an optimization problem on demand and using the initial parts of the generated solution to plan or control the system for the next few time steps. This technique is good for determining control inputs as it is aware of the dynamics of the modeled system, and thus can choose a control action that is optimal for not just the immediate state, but to also set up future states for success.
This package seeks to implement a MPC controller for a single track vehicle platform.
A basic kinematic bicycle model is used.
Receding horizon model predictive control is implemented.
When receiving a new trajectory, the next command is computed via a cold-started optimization problem. This means that the initial guess is the reference trajectory with no control inputs.
As the vehicle advances along a reference trajectory, previous solutions are rolled forward (warm starting), and reference points are back-filled from the refrence trajectory.
As the reference trajectory nears the end, reference weights are appropriately zeroed out to acheive the receding horizon behavior.
This controller uses the ControllerBase interface. As such, if the current state is "past" the current trajectory in some sense, a stopping deceleration command will be provided.
In all other cases, the library will solve an optimization problem and use the first element of the plan as the control command.
static_asserts are used as a coarse smell test for whether generated code has changed from the facade. They are generally placed around the code that might be affected.
A dynamic vehicle model should be used to more accurately represent the vehicle motion at higher velocities.
The model has been adapted from [1].
The following adaptations have been made:
These adaptations were made for the following reasons: