Information on maximum reward for roboschool environments?

[elitalobo]

No description provided.

[elitalobo]

Unable to find any documentation containing the baseline score or maximum reward details for roboschool environments.

[wpumacay]

Hi @elitalobo

I'm not sure if this is helpful, but you could take a look at the gym_forward_walker.py file. Inside the step method (line 119) the reward is computed from various components that make the whole reward function (+1 per alive-step, progress, actuation cost, ...). You could take a look at how they compute the rewards and get an idea of the actual ranges (I think the reward functions are not documented).

Perhaps what you are looking for is a standard range in which the rewards are going to be, like in the environments from dm_control. In their paper they specify that the rewards belong to a range of [0,1], and in their code they make the required scaling to make sure these ranges are in that range. For example, in the walker.py, they scale the components of the reward function to obtain a standard range of [0,1].