Unused parameter `state` passed in `enforceConstraints` function of `template NeuralNetModel`

[mithi]

In autorally/autorally_control/include/autorally_control/path_integral/neural_net_model.cut

• https://github.com/AutoRally/autorally/blob/kinetic-devel/autorally_control/include/autorally_control/path_integral/neural_net_model.cut

Lines 245 - 253, the MatrixXf &state is passed as a parameter to a function, but it's not used... How come? Thank you!

template<int S_DIM, int C_DIM, int K_DIM, int... layer_args>
void NeuralNetModel<S_DIM, C_DIM, K_DIM, layer_args...>::enforceConstraints(Eigen::MatrixXf &state, Eigen::MatrixXf &control)
{
  int i;
  for (i = 0; i < CONTROL_DIM; i++){
    if (control(i) < control_rngs_[i].x){
      control(i) = control_rngs_[i].x;
    }
    else if (control(i) > control_rngs_[i].y){
      control(i) = control_rngs_[i].y;
    }
  }
}

[gradyrw]

It's plausible that someone could write a new dynamics class which enforces both state and control constraints, so we wanted to accommodate for that case. This way, if someone does implement a dynamics class which enforces some state constraints, they don't need to modify any of the other path integral code.

[mithi]

@gradyrw

Thanks for clarifying.. I have this last two question before I close this issue:

In run_control_loop() From line 86 to 88 of: https://github.com/AutoRally/autorally/blob/kinetic-devel/autorally_control/include/autorally_control/path_integral/run_control_loop.cuh

    u = controller.computeControl(state); //Compute the control
    controller.model_->enforceConstraints(state, u);
    controller.model_->updateState(state, u); //Update the state using motion model.

But when you look at updateState() it also uses enforce constraints inside.. May I ask why this is necessary? updateState()

I want to clarify, is the current control parameter passed by reference u modified inside updateState() ?

In updateState() from: line 245 - 253 of: https://github.com/AutoRally/autorally/blob/kinetic-devel/autorally_control/include/autorally_control/path_integral/neural_net_model.cut

template<int S_DIM, int C_DIM, int K_DIM, int... layer_args>
void NeuralNetModel<S_DIM, C_DIM, K_DIM, layer_args...>::updateState(Eigen::MatrixXf &state, Eigen::MatrixXf &control)
{
  enforceConstraints(state, control);
  computeKinematics(state);
  computeDynamics(state, control);
  state += state_der_*dt_;
  state_der_ *= 0;
}

[gradyrw]

Yes, it looks like the enforceConstraints call before calling updateState is actually redundant in this case. This may be changed in the future, but for now, it is not doing any harm.

[JasonGibson274]

Closing since resolved