unilabsim/UniLab
A heterogeneous RL training framework for robotics that decouples CPU-based physics simulation (MuJoCo/MotrixSim) from GPU-based policy training on CUDA/ROCm.
Velocity · 7d
+5.2
★ / day
Trend
→steady
star history
UniLab enables robot reinforcement learning without relying on GPU-based simulation by leveraging a unified shared memory architecture between physics backends and neural network training. The system supports standard RL algorithms including PPO, SAC, and TD3 for training control policies on heterogeneous hardware. It provides simulation support via MuJoCo and MotrixSim while leveraging GPU resources for accelerated neural network optimization.