Use tanh squash

ml-agents/mlagents/trainers/ppo/trainer.py

             behavior_spec,
             self.trainer_settings,
             condition_sigma_on_obs=False,  # Faster training for PPO
+            tanh_squash=True,
             separate_critic=behavior_spec.action_spec.is_continuous(),
         )
         return policy

ml-agents/mlagents/trainers/torch/distributions.py

 
     def sample(self):
         sample = self.mean + torch.randn_like(self.mean) * self.std
-        return sample / 3
+        return sample
-        unscaled_val = value * 3  # Inverse of the clipping
-            -((unscaled_val - self.mean) ** 2) / (2 * var + EPSILON)
+            -((value - self.mean) ** 2) / (2 * var + EPSILON)
             - log_scale
             - math.log(math.sqrt(2 * math.pi))
         )