continuous comms

Project/Assets/ML-Agents/Examples/Hallway/Prefabs/SymbolFinderArea.prefab

     VectorObservationSize: 3
     NumStackedVectorObservations: 1
     m_ActionSpec:
-      m_NumContinuousActions: 0
-      BranchSizes: 0500000003000000
-    VectorActionSize: 0500000003000000
+      m_NumContinuousActions: 3
+      BranchSizes: 05000000
+    VectorActionSize: 
     VectorActionDescriptions: []
     VectorActionSpaceType: 0
     hasUpgradedBrainParametersWithActionSpec: 1
     VectorObservationSize: 3
     NumStackedVectorObservations: 1
     m_ActionSpec:
-      m_NumContinuousActions: 0
-      BranchSizes: 0500000003000000
-    VectorActionSize: 0500000003000000
+      m_NumContinuousActions: 3
+      BranchSizes: 05000000
+    VectorActionSize: 
     VectorActionDescriptions: []
     VectorActionSpaceType: 0
     hasUpgradedBrainParametersWithActionSpec: 1

Project/Assets/ML-Agents/Examples/Hallway/Scenes/HallwayCollab.unity

     m_EnableBakedLightmaps: 1
     m_EnableRealtimeLightmaps: 1
   m_LightmapEditorSettings:
-    serializedVersion: 10
+    serializedVersion: 12
     m_Resolution: 2
     m_BakeResolution: 40
     m_AtlasSize: 1024
     m_CompAOExponentDirect: 0
+    m_ExtractAmbientOcclusion: 0
     m_Padding: 2
     m_LightmapParameters: {fileID: 0}
     m_LightmapsBakeMode: 1
     m_PVRDirectSampleCount: 32
     m_PVRSampleCount: 500
     m_PVRBounces: 2
+    m_PVREnvironmentSampleCount: 500
+    m_PVREnvironmentReferencePointCount: 2048
+    m_PVRFilteringMode: 2
+    m_PVRDenoiserTypeDirect: 0
+    m_PVRDenoiserTypeIndirect: 0
+    m_PVRDenoiserTypeAO: 0
-    m_PVRFilteringMode: 1
+    m_PVREnvironmentMIS: 0
     m_PVRCulling: 1
     m_PVRFilteringGaussRadiusDirect: 1
     m_PVRFilteringGaussRadiusIndirect: 5
     m_PVRFilteringAtrousPositionSigmaAO: 1
-    m_ShowResolutionOverlay: 1
+    m_ExportTrainingData: 0
+    m_TrainingDataDestination: TrainingData
   m_LightingDataAsset: {fileID: 112000002, guid: 03723c7f910c3423aa1974f1b9ce8392,
     type: 2}
   m_UseShadowmask: 1
   m_GameObject: {fileID: 255077123}
   m_Enabled: 1
   m_EditorHideFlags: 0
-  m_Script: {fileID: 1077351063, guid: f70555f144d8491a825f0804e09c671c, type: 3}
+  m_Script: {fileID: 11500000, guid: 4f231c4fb786f3946a6b90b886c48677, type: 3}
   m_Name: 
   m_EditorClassIdentifier: 
   m_HorizontalAxis: Horizontal
   m_GameObject: {fileID: 255077123}
   m_Enabled: 1
   m_EditorHideFlags: 0
-  m_Script: {fileID: -619905303, guid: f70555f144d8491a825f0804e09c671c, type: 3}
+  m_Script: {fileID: 11500000, guid: 76c392e42b5098c458856cdf6ecaaaa1, type: 3}
   m_Name: 
   m_EditorClassIdentifier: 
   m_FirstSelected: {fileID: 0}
       propertyPath: m_Constraints
       value: 126
       objectReference: {fileID: 0}
-    - target: {fileID: 1410733827718496, guid: f3a451555dc514f46a69319857762eda, type: 3}
-      propertyPath: m_Name
-      value: symbol1Goal
-      objectReference: {fileID: 0}
     - target: {fileID: 1745841960385024, guid: f3a451555dc514f46a69319857762eda, type: 3}
       propertyPath: m_Name
       value: HallwayArea (1)
       propertyPath: m_RootOrder
       value: 5
       objectReference: {fileID: 0}
+    - target: {fileID: 1410733827718496, guid: f3a451555dc514f46a69319857762eda, type: 3}
+      propertyPath: m_Name
+      value: symbol1Goal
+      objectReference: {fileID: 0}
     m_RemovedComponents: []
   m_SourcePrefab: {fileID: 100100000, guid: f3a451555dc514f46a69319857762eda, type: 3}
 --- !u!1001 &812231986
   m_ClearFlags: 1
   m_BackGroundColor: {r: 0.46666667, g: 0.5647059, b: 0.60784316, a: 1}
   m_projectionMatrixMode: 1
+  m_GateFitMode: 2
+  m_FOVAxisMode: 0
-  m_GateFitMode: 2
   m_FocalLength: 50
   m_NormalizedViewPortRect:
     serializedVersion: 2
       objectReference: {fileID: 0}
     - target: {fileID: 4726744827719472, guid: f3a451555dc514f46a69319857762eda, type: 3}
       propertyPath: m_LocalEulerAnglesHint.z
+      value: 0
+      objectReference: {fileID: 0}
+    - target: {fileID: 4919701787480650, guid: f3a451555dc514f46a69319857762eda, type: 3}
+      propertyPath: m_LocalPosition.x
       value: 0
       objectReference: {fileID: 0}
     m_RemovedComponents: []

Project/Assets/ML-Agents/Examples/Hallway/Scripts/HallwayCollabAgent.cs

     public bool isSpotter = true;
     TextMesh m_MessageText;
     TextMesh m_MessageRec;
-    int m_Message = 0;
+    //int m_Message = 0;
+    float[] m_Message = new float[3];
 
 
     [HideInInspector]
 
     public override void OnEpisodeBegin()
     {
-        m_Message = -1;
+        //m_Message = -1;
+        System.Array.Clear(m_Message, 0, m_Message.Length);
         var agentOffset = 10f;
         if (isSpotter)
         {
         //{
         //    sensor.AddObservation(StepCount / (float)MaxStep);
         //}
-        sensor.AddObservation(toOnehot(m_Message));
+        //sensor.AddObservation(toOnehot(m_Message));
+        sensor.AddObservation(m_Message);
     }
 
     float[] toOnehot(int message)
         return onehot;
     }
 
-    public void tellAgent(int message)
+    public void tellAgent(float[] message)
+    //public void tellAgent(int message)
+    //{
+    //    m_Message = message;
+    //}
+
     public override void OnActionReceived(ActionBuffers actionBuffers)
     {
         AddReward(-1f / MaxStep);
         }
 
-        int comm_act = actionBuffers.DiscreteActions[1];
+        //int comm_act = actionBuffers.DiscreteActions[1];
+        float[] comm_act = actionBuffers.ContinuousActions.Array;
-            m_MessageText.text = "Sent:" + comm_act.ToString();
+            //m_MessageText.text = "Sent:" + comm_act.ToString();
+            m_MessageText.text = "Sent:" + comm_act[0].ToString("F2") + ", " + comm_act[1].ToString("F2") + ", " + comm_act[2].ToString("F2");
         }
         teammate.tellAgent(comm_act);
         // if (isSpotter) // Test

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

         )
         # this is a little bit of a hack but is whats recommended in the
         # gumbel softmax documentation
-        one_hot_diff_comms = obs[-1] - comms[1].detach() + comms[1]
-        obs[-1] = one_hot_diff_comms
+        #one_hot_diff_comms = obs[-1] - comms[1].detach() + comms[1]
+        #obs[-1] = one_hot_diff_comms
+        with torch.no_grad():
+            rand_n = (obs[-1] - comms.mean) / (comms.std + 1E-7)
+        obs[-1] = comms.mean + rand_n * comms.std
+
 
         log_probs, entropy, values = self.policy.evaluate_actions(
             obs,

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

         :return: Given the input, an AgentAction of the actions generated by the policy and the corresponding
         ActionLogProbs and entropies.
         """
-        comms = self._discrete_distribution.differentiable_forward(inputs, masks)
+        #comms = self._discrete_distribution.differentiable_forward(inputs, masks)
+        comms = self._continuous_distribution(inputs)
         return comms
 
     def forward(

Project/Assets/ML-Agents/Examples/Hallway/TFModels/diff-comms.onnx.meta

+fileFormatVersion: 2
+guid: 9d5d8d894cf434ee186ecd94a331ba3b
+ScriptedImporter:
+  internalIDToNameTable: []
+  externalObjects: {}
+  serializedVersion: 2
+  userData: 
+  assetBundleName: 
+  assetBundleVariant: 
+  script: {fileID: 11500000, guid: 683b6cb6d0a474744822c888b46772c9, type: 3}
+  optimizeModel: 1
+  forceArbitraryBatchSize: 1
+  treatErrorsAsWarnings: 0