modified clip_actions() to consider update frequency when dealing with acceleration

[zhum86]

I assume the expected behavior of clip_actions() would be modifying the action whenever the current action would make the next vehicle state (e.g., self.speed += action["acceleration"]*dt) outside its operation limit (e.g., max/min speed). The modified action should bring the vehicle state (e.g., speed) to its corresponding limit (e.g., max/min speed) in this step. The current implementation does not consider the dt (update frequency).

For example:

# Given
# self.MIN_SPEED = 0.0
# dt = 0.5 # s
# self.action["acceleration"] = -5.0
# self.speed = 0.5

def step(self, dt):
  self.clip_actions() # self.action["acceleration"] = -0.5 (current) vs -1.0 (new)
  ...
  self.speed += self.action["acceleration"] * dt # self.speed = 0.25 (current) vs 0.0 (new)

[zhum86]

@eleurent Hi, would you please review this PR?

[eleurent]

Hi, sorry for the late reply.

TLDR: this is intended behaviour to prevent abrupt deccelerations

I understand what you mean: this change makes sense if the goal is to ensure that the vehicle speed statisfies the hard constraints v_min <= v <= v_max by applying a correction in a single timestep dt.
However, the problem is that this can lead to very large accelerations (dv/dt) if dt is small.

The controller that is currently implemented instead follows a smoother trajectory, where any overflow from the maximum speed is corrected with a 1st-order controller with a response time of 1.0s, i.e it takes roughly 2 seconds to correct it.
I think this is slightly more physical, e.g. if a vehicle is initialised with a speed > v_max, it will slow down more gracefully that doing an instantaneous stop.

The drawback of that approach is that the constraint are not hard, since the correction is only applied after the speed is out of bounds, but I think that is fine as long as the overshoot stays small.