[Python] Low FPS (~2-3) When Running MuJoCo Simulation in LivelyBot Pi RL Baseline – Possible...

Intro Hi everyone,

I'm currently trying to reproduce the HighTorque-Robotics/livelybot_pi_rl_baseline project, which involves Sim2Sim reinforcement learning for a bipedal robot using both Isaac Gym and MuJoCo.

While Isaac Gym simulations run smoothly, I’m encountering a very low frame rate (~2-3 FPS) in MuJoCo, and I’m hoping someone here can help identify the root cause.

My setup Project Details:

Goal: Sim2Sim RL for LivelyBot using Isaac Gym + MuJoCo Hardware: Laptop with NVIDIA RTX 4080 GPU OS: Ubuntu 20.04 (NVIDIA drivers properly installed and active) MuJoCo Version: 2.3.6 Python Version: 3.8.20 Simulation Observations:

Isaac Gym: High GPU utilization, smooth performance. MuJoCo: ~2–3 FPS, extremely slow. GPU usage is negligible CPU usage is also low Troubleshooting Attempts:

Disabled matplotlib_thread → No improvement in FPS. Confirmed Isaac Gym works well → No hardware or PyTorch issues. Reduced resolution (e.g., 1280x720) → No noticeable improvement. MuJoCo performs well on other models Running MuJoCo’s humanoid.xml reaches 1000+ FPS. Tested LivelyBot model (pi_12dof_release_v1.xml) independently Using mj_step() manually for 5000 steps gives ~102 FPS. Viewer launched with mujoco.viewer.launch_passive()

My question ❓ Questions:

Why does MuJoCo perform so poorly (~3 FPS) in this project compared to Isaac Gym? Is there a known performance bottleneck when running MuJoCo with more complex robot models? Could it be related to physics parameters, viewer settings, or model configuration? Any recommended profiling tools or configuration tweaks to improve FPS in MuJoCo?

Continue reading...