Headings

Week 12: ML-Agents on M-Series Macs🛠 Phase 1: The Initial Setup🐍 Phase 2: Solving the Dependency Headache🛠 Phase 3: The Secret Fix🚀 Phase 4: Training & Implementation

Tags

fe
ml
ai
unity
fe
ml
ai
unity

Week 12: ML-Agents on M-Series Macs

J.Gong

2026-03-23

1.45min

Week 12: ML-Agents on M-Series Macs

Let’s be honest: setting up a Machine Learning environment on Apple Silicon (M1/M2/M3) can feel like a marathon. Recently, I’ve been diving into ML training, and doing it on a Mac Mini is… let’s say, an “adventure.” 🧗‍♂️

The silver lining? My Mac Mini is dead silent 🤫 compared to the jet-engine roar of a Windows gaming rig during training.

If you’re struggling to get Unity’s ML-Agents up and running, here is my battle-tested, step-by-step guide to making it actually work.

🛠 Phase 1: The Initial Setup

First, head over to the Official Unity ML-Agents Installation Guide.

Pro Tip: Follow the Advanced Installation path. You’ll need the source files for the dependencies to play nice later on.

Once installed, open the example files (found in the Project folder) in Unity. If you’re new to the workflow, check out the official sample guide on how to train a model.

🐍 Phase 2: Solving the Dependency Headache

When you first try to run the training command:

mlagents-learn config/ppo/3DBall.yaml --run-id=3DBall

…you’ll likely hit a wall because onnxscript is missing. 🛑

Go ahead and install the necessary dependencies. You might see some terrifying red text regarding version conflicts—just ignore them for now and push through:

python -m pip install onnx onnxscript onnxruntime

Next, you must downgrade/reinstall protobuf to a specific version, or the internal communication will fail:

python -m pip install protobuf==3.20.3

Verify the installation by running:

mlagents-learn --help

If you see the help menu, you’re halfway there! 🎉

🛠 Phase 3: The Secret Fix

Here is the part that drives everyone crazy. If you don’t do this, the .onnx file generated after training will be unreadable by Unity. 📄🚫

This happens because modern PyTorch uses a new exporter called Dynamo, while Unity’s importer expects the legacy format. You need to manually patch the ML-Agents library.

Locate this file in your environment: /opt/miniconda3/envs/mlagents/lib/python3.10/site-packages/mlagents/trainers/torch_entities/model_serialization.py

Find the torch.onnx.export call and add dynamo=False as shown below:

with exporting_to_onnx():
    torch.onnx.export(
        self.policy.actor,
        self.dummy_input,
        onnx_output_path,
        dynamo=False,  # <--- ADD THIS LINE HERE TO FIX UNITY COMPATIBILITY
        opset_version=SerializationSettings.onnx_opset,
        input_names=self.input_names,
        output_names=self.output_names,
        dynamic_axes=self.dynamic_axes,
    )
    logger.info(f"Exported {onnx_output_path}")

🚀 Phase 4: Training & Implementation

Now you’re ready!

  1. Run your training command. 🏃‍♂️💨
  2. Once finished, grab the generated .onnx file from the results folder.
  3. Drag it into your Unity project.
  4. Assign it to the Model field in the Behavior Parameters component of your Agent.

Happy training! 🎓🤖