Unity-Technologies/ml-agents · critical · UnityAgentsException

Can't use Behavior Type {behaviorType} without a model. Eith

Error message

Can't use Behavior Type {behaviorType} without a model. Either assign a model, or change to a different Behavior Type.

What it means

BehaviorParameters.GeneratePolicy() throws this when BehaviorType is set to InferenceOnly but no neural network model (m_Model) is assigned. Inference-only agents run entirely on a trained model, so without one there is no policy to create. The library demands either a model or a different behavior type that doesn't need one (e.g. HeuristicOnly).

Source

Thrown at com.unity.ml-agents/Runtime/Policies/BehaviorParameters.cs:240

        }

        void Awake()
        {
            OnPolicyUpdated += mode => { };
        }

        internal IPolicy GeneratePolicy(ActionSpec actionSpec, ActuatorManager actuatorManager)
        {
            switch (m_BehaviorType)
            {
                case BehaviorType.HeuristicOnly:
                    return new HeuristicPolicy(actuatorManager, actionSpec);
                case BehaviorType.InferenceOnly:
                    {
                        if (m_Model == null)
                        {
                            var behaviorType = BehaviorType.InferenceOnly.ToString();
                            throw new UnityAgentsException(
                                $"Can't use Behavior Type {behaviorType} without a model. " +
                                "Either assign a model, or change to a different Behavior Type."
                            );
                        }
                        return new SentisPolicy(actionSpec, actuatorManager, m_Model, m_InferenceDevice, m_BehaviorName, m_DeterministicInference);
                    }
                case BehaviorType.Default:
                    if (Academy.Instance.IsCommunicatorOn)
                    {
                        return new RemotePolicy(actionSpec, actuatorManager, FullyQualifiedBehaviorName);
                    }
                    if (m_Model != null)
                    {
                        return new SentisPolicy(actionSpec, actuatorManager, m_Model, m_InferenceDevice, m_BehaviorName, m_DeterministicInference);
                    }
                    else
                    {
                        return new HeuristicPolicy(actuatorManager, actionSpec);

View on GitHub (pinned to 3ecb446f75)

Solutions

Assign a trained model asset (.nn / ONNX) to the Model field of the BehaviorParameters component
Change Behavior Type to Heuristic Only or Default (which falls back to heuristic) if you don't have a model yet
Set the model programmatically via behaviorParameters.SetModel(behaviorName, model) before the Academy initializes
Verify the model asset import succeeded (Sentis/Unity Inference Engine imported the ONNX) so the field isn't silently null

Example fix

// before
behaviorParameters.behaviorType = BehaviorType.InferenceOnly; // Model left null
// after
behaviorParameters.behaviorType = BehaviorType.InferenceOnly;
behaviorParameters.model = trainedModelAsset; // assigned .nn/ONNX model

Defensive patterns

Strategy: validation

Validate before calling

// Unity C#: before Academy init / first step
if (behaviorParameters.BehaviorType == BehaviorType.InferenceOnly && behaviorParameters.Model == null)
{
    behaviorParameters.Model = trainedModelAsset; // or fall back to BehaviorType.HeuristicOnly
}

Try / catch

try
{
    academyInitialize();
}
catch (UnityAgentsException e) when (e.Message.Contains("without a model"))
{
    behaviorParameters.BehaviorType = BehaviorType.HeuristicOnly;
}

Prevention

Always assign the .nn model asset in the inspector before entering Play mode
Use BehaviorType.Default while iterating without a trained model
Validate BehaviorParameters in an editor script for all agents in the scene

When it happens

Trigger: Setting a BehaviorParameters component's Behavior Type to 'Inference Only' while the Model field is left empty; assigning a model at runtime via SetModel() with a null model; clearing the model in code or via prefab overrides before academy initialization.

Common situations: Dropping a trained NN model out of the inspector and forgetting to reassign; a training run failed so the .nn file was never produced; switching from Heuristic Only to Inference Only to test inference but never wiring up the model asset; prefab variants inheriting from a base that had no model.

Related errors

AI-assisted analysis of Unity-Technologies/ml-agents@3ecb446f75 (2026-09-02). Data as JSON: /api/errors/bd3dd850b34f53ab. Report an issue: GitHub.