· via dev.to (home feed)
China's Motphys claims MotrixSim trains robots 12.8x faster than Nvidia's Isaac Sim
Motphys says its Rust-built MotrixSim 1.0 engine reached a robot training target in 10.2 minutes versus 130.8 minutes for an Isaac Sim setup, though all figures come from the company's own tests.

What Motphys announced
Chinese company Motphys has unveiled MotrixSim 1.0, an embodied-intelligence simulation engine it says was developed fully in-house, and demonstrated it at CoRL 2026. According to Fathom News, which reported the announcement, Motphys said on Nov. 14 that it had been invited to the conference. The centerpiece claim: in the company's own tests, MotrixSim hit a training target roughly 12.8 times faster than a setup built on Nvidia's Isaac Sim.
Motphys positions the engine as a challenger in a field it says has long been dominated by Western players, naming Google's MuJoCo and Nvidia's Isaac Sim as the world's leading physics engines. Co-founder and CTO Zhang Heng put the ambition plainly: "Simulation once served to validate how machines perform. Now, it is where they learn how to perform."
The benchmark claims
Every performance figure in the announcement comes from Motphys itself. Fathom News notes this explicitly, and no independent verification exists so far.
The headline result comes from a G1 whole-body tracking task, where Motphys says both engines ran with an identical environment configuration and the same reinforcement learning algorithm. MotrixSim reportedly reached the current tracking score in about 10.2 minutes of end-to-end time, against 130.8 minutes for the Isaac Sim-based solution — a gap of about 12.8 times.
The company also published results from several physics tests:
- In a disturbance-rejection grasping test, where a Franka arm handled objects of different shapes, MotrixSim was the only engine to achieve a 100% pass rate at both 2 ms and 10 ms step sizes.
- In a critical friction test, the engine showed hold-and-slip behavior consistent with Coulomb's law and zero millimeters of slip distance.
- The Strandbeest mechanism, a structure with 92 linkages, stayed stable at a 5 ms step; Motphys says traditional engines fail to keep it stable and that MotrixSim preserves its assembly relationships during motion.
On throughput, Motphys says a pipeline combining CPU and GPU parallel computation lets the engine reach millions of physics steps per second across 4,096 parallel environments, with LiDAR simulation handling 124 million rays per second. The company also reports a 98.42% success rate on single-line solving and says visual synthesis throughput holds steady even when resolution rises 16-fold.
A Rust core and new tooling
Rather than layering on an existing open-source engine, Motphys says it wrote the MotrixSim core from scratch in Rust, arguing that the language delivers strong memory safety and computational power.
For developers, the company describes a "zero-conversion + AI auto-modeling" workflow and launched Forge, which it calls the industry's first Blender-based extension tool, allowing AI agents to automatically generate physics scenes. Motphys adds that its research has been accepted at conferences including RSS and CoRL, and that it has built a simulation and real-robot training center for embodied AI in China.
Why it matters
Reinforcement learning for robots depends on simulation: policies are trained across enormous numbers of trials before ever touching hardware, so engine speed directly sets the pace of research. A drop from roughly two hours to ten minutes per training run would let teams iterate far more within the same compute budget.
Nvidia's Isaac stack is deeply embedded in the robotics industry, and MuJoCo is the open-source standard across much of the research community, so a credible third option — particularly one designed for parallel scale from the start — would matter to anyone training embodied AI systems.
The caveat is that these are vendor-run benchmarks with no third-party replication, and the comparison setup was chosen by Motphys. The 12.8x figure should be read as a claim rather than a settled result. If it holds up under independent testing, MotrixSim becomes a serious entrant in robot simulation; until then, it is a striking data point from a company looking to break what it sees as a Western duopoly.
- #robotics
- #simulation
- #nvidia
- #embodied-ai
- #rust