additional changes

Project/Assets/ML-Agents/Examples/Bullet/Prefabs/BulletArea.prefab

     - target: {fileID: 5229809530250264056, guid: 7a515d80a9d5b4cfaba55aa066594016,
         type: 3}
       propertyPath: m_IsActive
-      value: 1
+      value: 0
       objectReference: {fileID: 0}
     m_RemovedComponents: []
   m_SourcePrefab: {fileID: 100100000, guid: 7a515d80a9d5b4cfaba55aa066594016, type: 3}
     - target: {fileID: 5229809530250264056, guid: 16266ae2c040142b1a996878d94ec3fb,
         type: 3}
       propertyPath: m_IsActive
-      value: 0
+      value: 1
       objectReference: {fileID: 0}
     m_RemovedComponents:
     - {fileID: 9204493211719777360, guid: 16266ae2c040142b1a996878d94ec3fb, type: 3}

Project/Assets/ML-Agents/Examples/Bullet/Scripts/AttentionSensorComponent.cs

         var bullets = m_AgentTransform.parent.GetComponentsInChildren<Bullet>();
         // Sort by closest :
         Array.Sort(bullets , (a, b) => Vector3.Distance(a.transform.position, m_AgentTransform.position) - Vector3.Distance(b.transform.position, m_AgentTransform.position) > 0 ? 1 : -1);
+
-            m_ObservationBuffer[m_CurrentNumObservables * m_ObservableSize + 0] = b.transform.position.x - m_AgentTransform.position.x;
-            m_ObservationBuffer[m_CurrentNumObservables * m_ObservableSize + 1] = b.transform.position.z - m_AgentTransform.position.z;
+            m_ObservationBuffer[m_CurrentNumObservables * m_ObservableSize + 0] = (b.transform.position.x - m_AgentTransform.parent.position.x) / 10f;
+            m_ObservationBuffer[m_CurrentNumObservables * m_ObservableSize + 1] = (b.transform.position.z - m_AgentTransform.parent.position.z) / 10f;
             m_ObservationBuffer[m_CurrentNumObservables * m_ObservableSize + 2] = b.transform.forward.x;
             m_ObservationBuffer[m_CurrentNumObservables * m_ObservableSize + 3] = b.transform.forward.z;
             m_CurrentNumObservables += 1;

Project/Assets/ML-Agents/Examples/Bullet/Scripts/DodgeAgent.cs

     public override void CollectObservations(VectorSensor sensor)
     {
 
-        sensor.AddObservation(transform.position.x - area.transform.position.x);
-        sensor.AddObservation(transform.position.z - area.transform.position.z);
+        sensor.AddObservation((transform.position.x - area.transform.position.x) / 10f);
+        sensor.AddObservation((transform.position.z - area.transform.position.z) / 10f);
 
     }
     /// <summary>

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

             normalize=self.normalize,
         )
 
-
         total_enc_size = encoder_input_size + encoded_act_size
 
         self.self_embedding = LinearEncoder(6, 1, 64)
 
         # TODO : This is a Hack
         var_len_input = vis_inputs[0].reshape(-1, 20, 4)
-        key_mask = 0 * (
+        key_mask =  (
             torch.sum(var_len_input ** 2, axis=2) < 0.01
         )  # 1 means mask and 0 means let though
 
-        objects = torch.cat([expanded_x_self, var_len_input], dim=2)  #(b,20,68)
+        objects = torch.cat([expanded_x_self, var_len_input], dim=2)  # (b,20,68)
-        obj_encoding = self.obs_embeding(objects)#(b,20,64)
+        obj_encoding = self.obs_embeding(objects)  # (b,20,64)
-        self_and_key_emb = torch.cat([x_self.reshape(-1,1,64), obj_encoding], dim=1) #(b,21,64)
+        self_and_key_emb = torch.cat(
+            [x_self.reshape(-1, 1, 64), obj_encoding], dim=1
+        )  # (b,21,64)
-        ) # first one is never masked
+        )  # first one is never masked
 
         output, _ = self.attention(
             self_and_key_emb, self_and_key_emb, self_and_key_emb, key_mask
-
-        ) / torch.sum(1 - key_mask, dim=1, keepdim=True) # average pooling
-
-
+        ) / torch.sum(
+            1 - key_mask, dim=1, keepdim=True
+        )  # average pooling
-
 
         if self.use_lstm:
             # Resize to (batch, sequence length, encoding size)