Motphys says MotrixSim 1.0 is 12.8x faster than Isaac Sim on G1 task

BEIJING -- November 19, 2026 -- Chinese simulation developer Motphys says its MotrixSim 1.0 physics engine reaches a target G1 whole-body tracking score in about 10.2 minutes of end-to-end time. An Isaac Sim-based solution needs 130.8 minutes for the same score. By our calculation, that is a gap of about 12.8 times, or a cut of roughly 92% (120.6 minutes).

Motphys showed the engine at CoRL 2026. The company says both runs used the same task environment configuration and the same reinforcement-learning algorithm. It did not disclose the hardware behind either run.

Motphys reports a 100% pass rate in Franka grasping tests at two step sizes

In a disturbance-rejection grasping test with a Franka arm and objects of different shapes, the company says MotrixSim 1.0 was the only engine to pass every trial at both 2 ms and 10 ms simulation step sizes. In a critical-friction test, it reports hold-and-slip behavior consistent with Coulomb's law and a slip distance of 0 mm.

A third test used a Strandbeest mechanism with 92 linkages at a 5 ms time step. Motphys says traditional engines fail to keep the mechanism stable at that step. It says MotrixSim solves the closed-loop constraints together with the multibody dynamics in generalized coordinates, so the assembly relationships hold during motion.

Motphys wrote the engine core in Rust instead of building on open-source engines

The company says it developed the core from the ground up in Rust, which it credits with memory safety and computation power. The release names Google's MuJoCo and NVIDIA's Isaac Sim as the leading engines today.

Motphys says a CPU/GPU parallel pipeline delivers millions of physics steps per second across 4,096 parallel environments. It also reports LiDAR simulation throughput of 124 million rays per second. Visual synthesis, it says, keeps stable throughput when resolution rises 16 times.

On workflow, Motphys cites a 98.42% success rate for high-speed solving in a single line of code. It also describes a "zero-conversion + AI auto-modeling" workflow. It launched Forge, a Blender-based extension that lets AI agents generate physics scenes automatically.

Shorter training loops change iteration time for robot developers

The G1 result converts to about 2.2 hours (130.8 minutes) on the Isaac Sim-based setup versus about 10 minutes on MotrixSim. On the company's figures, roughly 12 full MotrixSim runs fit into the time of one Isaac Sim-based run. The comparison covers a single task, and the release gives no timings for other workloads.

"Simulation once served to validate how machines perform. Now, it is where they learn how to perform," said Zhang Heng, co-founder and CTO of Motphys.

Motphys says its research has been accepted at RSS and CoRL. It has also built a simulation and real-robot training center for embodied AI in China.

Published by:
fairsonline_team
Products
Target market(s)