Moore Threads Open-Sources MuJoCo Warp MUSA for End-to-End Embodied AI
Chinese GPU maker Moore Threads announced it has established a complete end-to-end pipeline for embodied AI—covering synthetic data generation, large model training, physics simulation verification, and edge hardware deployment. The company also open-sourced MuJoCo Warp MUSA, a physics simulation backend adapted for its proprietary MUSA GPU architecture. This development demonstrates how full-function domestic GPUs can meet the combined computational demands of AI training, physics simulation, graphics rendering, and edge inference. It offers domestic robotics developers an alternative to Western hardware platforms like NVIDIA CUDA for Sim2Real training. Moore Threads highlighted its full-stack simulation platform, MT Lambda, and confirmed successful Sim2Real hardware verification on quadruped robot dogs and bipedal humanoid robots. Additionally, the company established an industrial embodied AI innovation center in Wuxi and formed a joint laboratory with China's national embodied AI mid-test base.
## BACKGROUND
Embodied AI refers to artificial intelligence integrated into physical platforms like robots, enabling them to interact dynamically with the physical world. MuJoCo Warp is a GPU-accelerated physics simulation framework originally created for fast robotic policy training on NVIDIA CUDA platforms. Sim2Real is the process of transferring control policies trained inside virtual physics simulators onto physical hardware in the real world.