Accelerating Mujoco Simulations in Unity

[tkwk9]

Intro

Hello! I am a Unity hobbyist learning about Mujoco for Unity.

My setup

I am using the Mujoco for Unity version 3.3.0. I am developing on M2 Macbook, running on Sequoia 15.3.1

My question

I'm using Mujoco for Unity to run physics simulations in my project, where a model is trained within a Mujoco physics environment. To speed up training, I needed a way to advance the simulation manually rather than relying on Unity’s standard FixedUpdate. I couldn’t find a built-in solution, so I modified MjScene.cs as follows:

• Commented out the original MjScene::FixedUpdate method.
• Added a new public method MjScene::ManualStep that replicates the functionality of the original FixedUpdate.

# MjScene.cs

protected unsafe void FixedUpdate() {
  // preUpdateEvent?.Invoke(this, new MjStepArgs(Model, Data));
  // StepScene();
  // postUpdateEvent?.Invoke(this, new MjStepArgs(Model, Data));
}

public unsafe void ManualStep() {
  preUpdateEvent?.Invoke(this, new MjStepArgs(Model, Data));
  StepScene();
  postUpdateEvent?.Invoke(this, new MjStepArgs(Model, Data));
}

I then do the following in a separate class:

# TrainingManager.cs

// Training Simulation (Coroutine-Based)
IEnumerator TrainingLoop() {
    while (!terminated) {
        // Manually step the simulation
        MjScene.Instance.ManualStep();

        var observations = environment.GetObservations();
        if (environment.ShouldTerminate(observations))
            environment.Reset();
            break;

        environment.UpdateReward(observations);
        var actions = agent.GetActions(observations);
        environment.ApplyActions(actions);

        // In the real code, I yield ~30 times per second.
        yield return null; 
    }
}

// "Real" Simulation (FixedUpdate-Based)
void FixedUpdate() {
    // Skip FixedUpdate while training
    if (isTraining || agent == null) return;

    // Manually step the simulation (same as in training)
    MjScene.Instance.ManualStep();

    var observations = environment.GetObservations();
    if (environment.ShouldTerminate(observations)) {
        environment.Reset();
        return;
    }

    environment.UpdateReward(observations);
    var actions = agent.GetActions(observations);
    environment.ApplyActions(actions);
}

I don't modify Time.fixedDeltaTime, and my understanding is that the plugin relies on this value to determine the simulation step interval in Mujoco.

Problem

Even though both the training and real environments use the same ManualStep implementation to advance the simulation, the results produced are consistent within each mode but differ between them.

Questions

• Is there a more standard way to accelerate Mujoco simulations for training in Unity without modifying MjScene.cs?
• What could be causing the discrepancy in simulation outcomes between the coroutine-based training loop and the FixedUpdate-driven "real" simulation?

Minimal model and/or code that explain my question

No response

Confirmations

• I searched the latest documentation thoroughly before posting.
• I searched previous Issues and Discussions, I am certain this has not been raised before.

[tkwk9]

Hi @Balint-H , could I get your thoughts on this?

[Balint-H]

The first thing that comes to mind is upping the time scale (essentially calls FixedUpdate more frequently than what dT would indicate). For single-environment scenes 2 or 3 times speed should be managed easily on most systems and models.

Let me think about what might cause the coroutine method to have different results. Does any non-mujoco script or component impact the observations/rewards?

[tkwk9]

Changing the timescale seems to work – thank you so much! It looks like I may have overcomplicated things a bit.. It still feels strange to me that the coroutine approach didn’t work – it seems like it should have. But my problem is solved, I think – thank you again!

Just to answer the Does any non-mujoco script or component impact the observations/rewards?:

I don't believe so. The scene is purely in mujoco, and I set the Unity simulation mode to SimulationMode.Script during debugging to make sure to disable any Unity physics.