Question about the action normalization

[Beanpow]

Hi, I found that the following action normalization is useful to train an agent. Could you please provide additional details on how the normalization is derived? For example, how do you determine the coefficient of the sigmoid-like function? Thanks in advance.

$a = \frac{1}{1+e^{-5(a-0.5)}}$

myosuite/myosuite/envs/myo/base_v0.py

Lines 80 to 84 in 4d146e1

	# Explicitely project normalized space (-1,1) to actuator space (0,1) if muscles
	if self.sim.model.na and self.normalize_act:
	# find muscle actuators
	muscle_act_ind = self.sim.model.actuator_dyntype==3
	muscle_a[muscle_act_ind] = 1.0/(1.0+np.exp(-5.0*(muscle_a[muscle_act_ind]-0.5)))

[Vittorio-Caggiano]

hi @Beanpow ,
sorry, I missed this issue. In the very beginning of this project we finetuned the normalization to be stable with a range of RL explorations. We also provide the non-normalized option. Please, let us know if there are cases in which this normalization doesn't work