diff --git a/torchy_baselines/sac/policies.py b/torchy_baselines/sac/policies.py
index 61080c0..1cc6fdf 100644
--- a/torchy_baselines/sac/policies.py
+++ b/torchy_baselines/sac/policies.py
@@ -51,7 +51,7 @@ class Actor(BaseNetwork):
             # Avoid saturation by limiting the mean of the gaussian to be in [-1, 1]
             # self.mu = nn.Sequential(self.mu, nn.Tanh())
             # TODO: test with small positive slope to have non zero gradient
-            self.mu = nn.Sequential(self.mu, nn.Hardtanh(min_val=-2.0, max_val=2.0))
+            self.mu = nn.Sequential(self.mu, nn.Hardtanh(min_val=-3.0, max_val=3.0))
         else:
             self.action_dist = SquashedDiagGaussianDistribution(action_dim)
             self.mu = nn.Linear(net_arch[-1], action_dim)