An engine written in-house
Motphys, a Chinese company, presented MotrixSim 1.0 at CoRL 2026, describing it as an embodied intelligence simulation engine built entirely by its own team. The company said the core was written from the ground up in Rust rather than on top of existing open-source engines, giving it memory safety and computation power. Zhang Heng, co-founder and CTO, said simulation had shifted from checking how a machine performs to being the place where it learns to perform.
Only engine at 100%
In a disturbance-rejection grasping test in which a Franka arm picks up objects of different shapes, MotrixSim 1.0 was the only engine to reach a 100% pass rate at simulation step sizes of both 2 ms and 10 ms, the company said. In a test at the critical friction point, it reproduced hold-and-slip behaviour consistent with Coulomb's law, with a slip distance of 0 mm.
Training time against Isaac Sim
MotrixSim runs a CPU and GPU parallel pipeline that the company says reaches a physics throughput of millions of steps per second across 4,096 parallel environments. On a G1 whole-body tracking task, run with the same task configuration and the same reinforcement learning algorithm, MotrixSim reached its tracking score in about 10.2 minutes of end-to-end time, while an Isaac Sim-based solution took 130.8 minutes, a gap of about 12.8 times.
Other figures and a Blender tool
For a Strandbeest closed-loop mechanism with 92 linkages, MotrixSim held the assembly relationships at a 5 ms time step, which the company says traditional engines failed to do, by solving closed-loop constraints together with multibody dynamics in generalized coordinates. LiDAR simulation throughput reaches 124 million rays per second, and visual synthesis keeps its throughput when resolution rises 16 times. Motphys also launched Forge, described as a Blender-based extension tool that lets AI agents generate physics scenes, and cited single-line solving of 98.42% success. It said its research has been accepted at RSS and CoRL and that it has built a simulation and real-robot training centre for embodied AI in China. The release names Google's MuJoCo and NVIDIA's Isaac Sim among the leading physics engines worldwide.
