[Proposal] only positive reward

[fan-ziqi]

Proposal

Users can selectively calculate only positive rewards in reward manager

[mpgussert]

Users have full control over the reward function when defining their own environment. It's not clear to me what you are asking for here. Do you want a toggle to invert the rewards on the stand alone samples?

[fan-ziqi]

Thank you for reply! @mpgussert

I mean to say, the compute_reward function in legged_gym uses only_positive_rewards to control whether the reward is reduced to 0. Do you plan to introduce this feature in reward_manager as well?

    def compute_reward(self):
        """ Compute rewards
            Calls each reward function which had a non-zero scale (processed in self._prepare_reward_function())
            adds each terms to the episode sums and to the total reward
        """
        self.rew_buf[:] = 0.
        for i in range(len(self.reward_functions)):
            name = self.reward_names[i]
            rew = self.reward_functions[i]() * self.reward_scales[name]
            self.rew_buf += rew
            self.episode_sums[name] += rew
        if self.cfg.rewards.only_positive_rewards:
            self.rew_buf[:] = torch.clip(self.rew_buf[:], min=0.)
        # add termination reward after clipping
        if "termination" in self.reward_scales:
            rew = self._reward_termination() * self.reward_scales["termination"]
            self.rew_buf += rew
            self.episode_sums["termination"] += rew