URLab_bridge example policy performance

[mulegan]

Summary

ZMQ Network Connection Established Successfully, but Robot in UE5 Does Not Respond to Control Commands

Environment

• Unreal Engine version:5.7
• OS (Windows / Linux) and version:Windows+WSL
• URLab plugin version / commit SHA:f5e02af2ac021e679aee4021593b38c41b77c5b2
• Visual Studio version (Windows only):2022

Reproduction steps

1. Launch Unreal Side: Open the project in UE5, load a level containing an AMjManagerand the robot Blueprint, and press Play​ to start the simulation.
2. Launch Python Side: In a terminal, run the command uv run src/run.py --ui. In the Dashboard GUI, select the target robot (e.g., g1) and a policy (e.g., unitree_12dofor beyondmimic).
3. Observe Result: The Dashboard logs indicate a successful ZMQ connection and the policy begins computing and sending commands, but the robot's joints in the Unreal viewport show no movement.

Expected behaviour

When control commands are sent from the Python side via ZMQ, the corresponding joints of the robot in UE5 should move accordingly.

Actual behaviour

torso=[0.050,-0.237,0.139]
BeyondmimicCtrl dance1_subject2: 0%| 16:01:31 urlab_policy.dashboard INFO — [DIAG step=41250] ts=41300 action_rms=2.856 dof_err=2.619 base_pos=[0.023,-0.210,0.117] base_quat=[0.268,0.460,-0.212,0.820] lin_vel=[0.000,0.000,-0.000] ang_vel=[0.000,-0.000,-0.000] torso=[0.050,-0.237,0.139]
BeyondmimicCtrl dance1_subject2: 0%| 16:01:36 urlab_policy.dashboard INFO — [DIAG step=41500] ts=41550 action_rms=2.856 dof_err=2.619 base_pos=[0.023,-0.210,0.117] base_quat=[0.268,0.460,-0.212,0.820] lin_vel=[0.000,0.000,-0.000] ang_vel=[0.000,-0.000,-0.000] torso=[0.050,-0.237,0.139]
BeyondmimicCtrl dance1_subject2: 0%| 16:01:41 urlab_policy.dashboard INFO — [DIAG step=41750] ts=41800 action_rms=2.856 dof_err=2.619 base_pos=[0.023,-0.210,0.117] base_quat=[0.268,0.460,-0.212,0.820] lin_vel=[0.000,0.000,-0.000] ang_vel=[0.000,-0.000,-0.000] torso=[0.050,-0.237,0.139]
BeyondmimicCtrl dance1_subject2: 0%| | 0/6574 [14:40<?, ?step/s]
Beyondmimic Jump_wose: 42019.0step [14:40, 49.45step/s]

Relevant logs

Before submitting

• I searched existing issues and didn't find a duplicate.
• I'm on a supported Unreal Engine version (5.7 or newer), or I've noted why not in the environment block.
• I reproduced this on the latest main (or noted the commit I'm on).

[mulegan]

Thank you for the previous guidance. I've resolved the initial issue by adding a UMjPDControllercomponent to the target articulation (g1_29dof_rev_1_0_C_1) in the Unreal level, and the ZMQ control pipeline is now active. The robot's joints are receiving and responding to the control commands.
However, a new problem has emerged: when running the BeyondMimicDance​ policy, the robot cannot maintain stability. It immediately collapses, falls to the ground, and appears to struggle​ instead of performing the dance motions.

[mulegan]

Great news – I've resolved the initial connectivity issue! I can now successfully use the project's BeyondMimicDance​ demo to drive the G1 robot in UE5.7. The control pipeline is active, and the robot responds to the dance commands.
However, a new challenge has emerged: the robot consistently loses balance and falls over​ during the dance sequence. I'm currently unsure how to address this stability problem.
To summarize the current situation:
✅ Connection & Control Established: The robot (g1_29dof_rev_1_0_C_1) has a UMjPDControllercomponent, and the ZMQ bridge is operational.
❌ Stability Issue: The robot cannot maintain equilibrium while executing the BeyondMimicDance motions and collapses.
What I've tried/checked:

• Adjusted the UMjPDControllergains (Kp/Kv) to increase joint stiffness.
• Verified ground collision and gravity settings.

My main questions are:

1. What are the recommended approaches or best practices within the UnrealRoboticsLab framework to improve whole-body stability​ for dynamic motions like dancing?
2. Are there specific configurations for the UMjPDController, the policy, or the simulation environment that are crucial for balancing humanoid robots like the G1?
3. Could the issue be related to the initial pose​ of the robot or the motion data​ itself?

Any guidance on how to proceed would be greatly appreciated. Thank you!

[mulegan]

❗️After successfully establishing the ZMQ connection, the robot can be driven by the policy but cannot execute motions stably. Specifically: when using the BeyondMimicDancepolicy, the robot’s joints respond, but it quickly loses balance and falls, failing to perform the dance motion.
Step：

1. Start Unreal Simulation: Open the project in UE5 Editor, load the level containing the AMjManagerand the robot g1_29dof_rev_1_0_C_1, and press Play.
2. Start Python Client: In a terminal, run uv run src/run.py --ui.
3. Select and Run Policy: In the Dashboard GUI, select the robot prefix g1_29dof_rev_1_0_C_1and the policy BeyondMimicDance, then click Start Policy.
4. Observe Result: The robot’s joints begin to move but lose balance within 14-30 seconds and fall to the ground.

[jonathanembleyriches]

❗️After successfully establishing the ZMQ connection, the robot can be driven by the policy but cannot execute motions stably. Specifically: when using the BeyondMimicDancepolicy, the robot’s joints respond, but it quickly loses balance and falls, failing to perform the dance motion. Step：

1. Start Unreal Simulation: Open the project in UE5 Editor, load the level containing the AMjManagerand the robot g1_29dof_rev_1_0_C_1, and press Play.
2. Start Python Client: In a terminal, run uv run src/run.py --ui.
3. Select and Run Policy: In the Dashboard GUI, select the robot prefix g1_29dof_rev_1_0_C_1and the policy BeyondMimicDance, then click Start Policy.
4. Observe Result: The robot’s joints begin to move but lose balance within 14-30 seconds and fall to the ground.

These policies that we have provided examples for are pre-trained policies not made or trained by us. We have only included them as an example. So while we can provide some advice they are not guaranteed to be perfect.

To get the best performance out of them you may have to tweak the physics timestep and other parameters (please see this comment and the thread for some general tips and tricks. More specifically it is worth looking at the "SimOptions" to tweak until you get the best performance.

For the release video we managed to tweak the time step (and above recommendations) to get the dance, violin, and AMO locomotion to perform somewhat robustly on flat ground.

My main questions are:

1. What are the recommended approaches or best practices within the UnrealRoboticsLab framework to improve whole-body stability​ for dynamic motions like dancing?
2. Are there specific configurations for the UMjPDController, the policy, or the simulation environment that are crucial for balancing humanoid robots like the G1?
3. Could the issue be related to the initial pose​ of the robot or the motion data​ itself?

Any guidance on how to proceed would be greatly appreciated. Thank you!

To answer your specific questions:

1. See above (the editor cpu throttling is a good change to make regardless)
2. The PD controller should receive all of its gains over ZMQ by the urlab_bridge itself. so tweaking this shouldn't be necessary.
3. I believe the reason the dance policies are not super robust is because the policy is trying to match the motion data and as soon as there is a bit too much drift the robustness falls apart. (If you look into our approach I believe we use IK on the policy side).

If may we worth taking a look at the quadruped WTW controller which is a lot more robust.

While not currently a priority on my roadmap until after alpha, I do want to have a library of OpenSource controllers that URLab_bridge can easily run with more guaranteed robustness. The plan is that users can just grab and go an xml and a policy and have it working (or we even provide the raw .uasset). A start would be to use policies trained within the MuJoCo eco system (mjlab, etc) as this would give the best sim-to-sim results.

This was the main goal behind the URLab_bridge to try and build this API out as well as being the middleware for external connections. If you would be interested in discussing this lets continue this on the bridge repo here!

[mulegan]

Currently, if I start the UE5.7 Runtime first and then launch the controller from the URLab_bridge project, it hangs. I have to restart the UE5.7 Runtime mode to get it to drive properly. How can this issue be resolved, or has it already been fixed in a recent commit?

[jonathanembleyriches]

Currently, if I start the UE5.7 Runtime first and then launch the controller from the URLab_bridge project, it hangs. I have to restart the UE5.7 Runtime mode to get it to drive properly. How can this issue be resolved, or has it already been fixed in a recent commit?

I'm not sure what process you mean is hanging, is it the sim or bridge - what is freezing? Can you please provide some more context. A video would be great.

If it is the urlab_bridge hanging then please also paste the logs.

[mulegan]

Sorry, I tried to upload a video but it kept failing. Not sure if it's a setup issue.
Let me rephrase my problem: I feel there might be an issue with the steps I'm using for the BeyondMimic strategy. I'm not sure if it's already fixed or if I haven't updated to the latest commit. My process is:

1. Start UE5 and urlab_bridge first.
2. Start UE5 Runtime, then urlab_bridge can detect the available robot models. After that, I select the corresponding robot and driving policy, but at this step, it gets stuck.
3. I have to restart UE5 Runtime to proceed to the next step.

I'm not sure what's causing this issue. Below are screenshots of the process.
​​