Open twu-mrc opened 2 years ago
https://github.com/leggedrobotics/rsl_rl/blob/master/rsl_rl/modules/actor_critic.py#L121 The action noise had better use positive parameterization, e.g. self.distribution = Normal(mean, torch.ones_like(mean) + nn.functional.elu(self.std - 1))
According to @twu-mrc, self.distribution in rsl_rl 1.0.2 version is L121 and 2.0.2 version is L101, in _updatedistribution() function. The self.distribution below the comment 'Action noise' is None.
https://github.com/leggedrobotics/rsl_rl/blob/master/rsl_rl/modules/actor_critic.py#L121 The action noise had better use positive parameterization, e.g. self.distribution = Normal(mean, torch.ones_like(mean) + nn.functional.elu(self.std - 1))