allow tuple action space (#11429)

Co-authored-by: Jiajie Xiao <jj@Jiajies-MBP-2.attlocal.net>

rllib/models/tf/recurrent_net.py

@@ -113,7 +113,10 @@ class LSTMWrapper(RecurrentNetwork):
         self.cell_size = model_config["lstm_cell_size"]
         self.use_prev_action_reward = model_config[
             "lstm_use_prev_action_reward"]
-        self.action_dim = int(np.product(action_space.shape))
+        if action_space.shape is not None:
+            self.action_dim = int(np.product(action_space.shape))
+        else:
+            self.action_dim = int(len(action_space))
         # Add prev-action/reward nodes to input to LSTM.
         if self.use_prev_action_reward:
             self.num_outputs += 1 + self.action_dim

rllib/models/torch/recurrent_net.py

@@ -114,7 +114,10 @@ class LSTMWrapper(RecurrentNetwork, nn.Module):
         self.time_major = model_config.get("_time_major", False)
         self.use_prev_action_reward = model_config[
             "lstm_use_prev_action_reward"]
-        self.action_dim = int(np.product(action_space.shape))
+        if action_space.shape is not None:
+            self.action_dim = int(np.product(action_space.shape))
+        else:
+            self.action_dim = int(len(action_space))
         # Add prev-action/reward nodes to input to LSTM.
         if self.use_prev_action_reward:
             self.num_outputs += 1 + self.action_dim