Motphys Series A

Motphys Deep Dive: From a Game Action-Physics Engine to an Embodied-AI Simulation Substrate — Building the 'Virtual Training Ground' for Physical AI with the Fully In-House MotrixSim Engine

Motphys (Beijing Mouxianfei Technology) is a domestically built embodied-simulation infrastructure provider. Its fully in-house MotrixSim physics engine anchors a stack spanning MotrixLab training and MotrixGen data synthesis, pursuing a Real2Sim2Real loop for zero-shot sim-to-real transfer. Sequoia China, Hillhouse and MiraclePlus backed the angel round; Huawei Hubble invested strategically in 2024.

Executive Summary

Motphys (Beijing Mouxianfei Technology Co., Ltd.) is a Chinese provider of embodied-AI simulation infrastructure. Founded around 2021, it develops a full-stack, proprietary simulation platform built around its own real-time physics engine, MotrixSim, and pursues a "Real2Sim2Real" technical path: reconstruct real scenes in simulation, train policies at massive parallelism, then transfer them to physical robots with zero shots. Its product stack layers the MotrixLab imitation-and-reinforcement-learning framework and the MotrixGen synthetic-data pipeline above the engine, supplemented by MotrixStudio, an asset/scene editor, and a SimReady asset system. The company positions itself as a "virtual training ground" for robots, spanning locomotion, dexterous manipulation and, since 2026, automotive and industrial machinery.

Motphys is distinguished by two things. First, it insists on building the lowest-level physics engine in-house — from the math library up, with no external physics-library dependency — using Rust for performance and memory safety, generalized-coordinate whole-system solving, and an LCP contact/friction solver. Second, it has secured an unusually strong domestic backing for a simulation startup: Sequoia China Seed, Hillhouse Ventures and MiraclePlus at the angel round, and a strategic investment from Huawei's Hubble in 2024, alongside a joint "embodied simulation base" solution unveiled with Huawei in September 2026.

This report treats Motphys as an infrastructure-layer company whose strategic value lies in the efficiency ceiling it sets for downstream training and in the degree of domestic control it offers over a stack long dominated by NVIDIA Isaac. It scores the company 7.5 on technology, 6.5 on business, 6.5 on financing, 7.5 on product and 7.5 on market, for a composite of 7.1. The discount relative to its technical ambition reflects a single unresolved gap: published, independently verifiable revenue and signed customer contracts remain thin, and several headline benchmark and financing claims come from the company or its own press materials.

Company Overview

Public records and company materials place Motphys's founding around 2021, with the Beijing entity registered in Haidian District. There is a persistent discrepancy between sources over the exact date and legal entity: the company website and some databases give 9 August 2021 for "Beijing Mouxianfei Technology Co., Ltd.", while 2022-era media and industry profiles describe "Beijing Mouxianfei Technology Co., Ltd." as founded in October 2020. The two dates may reflect an earlier operating entity versus a later re-registration, a point this report records as unresolved in DISCREPANCIES.

Founder and CEO Cui Hanqing is a Wuhan University computer-science alumnus and a 2007 ACM-ICPC Nanjing regional gold medalist — the first such gold in Wuhan University's history. He interned at Microsoft Research Asia, worked at Microsoft China for two and a half years, then became a technical lead and later a partner at Keystone (Gongding Shi), a sports-game company backed by Sequoia. He founded Motphys after that, pursuing an MBA at Tsinghua University, where he met key team members. The core team reportedly includes five ACM gold medalists and graduates of Tsinghua, Peking University, Wuhan University, Harbin Institute of Technology, Cambridge, Cornell and Waseda. Co-founder and CTO Zhang Heng is the public face of the company's technical roadmap, presenting the MotrixSim 1.0 launch in November 2026.

Motphys operates a Beijing headquarters, a Hangzhou R&D center and a Wuhan training center. In October 2025 it signed an agreement with the Wuhan East-Lake (Dongxihu) / Tianhe Airport economic zone to build what both sides call China's first robot simulation-training ground with a fully in-house stack from physics engine to data pipeline; the company is setting up a central-China headquarters and R&D center there. The founding team's Wuhan connection is personal: Cui studied at Wuhan University and has described the return as a homecoming. Headcount is reported in the 51–100 range by third-party databases; the company has not confirmed a current figure.

Product and Technology

MotrixSim is the base layer: a proprietary real-time physics engine that the company says is written from scratch in Rust with zero external physics-library dependency. It emphasizes whole-system dynamics under generalized coordinates rather than solving bodies independently and patching constraints afterward, plus an LCP-based solver for normal support and tangential friction at contacts. The engine targets the failure modes that matter for manipulation: stable multi-body contact, rolling at critical friction, joint accuracy below 0.1% PD-control error against analytical solutions, and conservation behavior illustrated by Newton's-cradle and gyroscopic-precession tests. Rendering uses PBR, screen-space global illumination (SSGI), image-based lighting (IBL) and a hybrid 3DGS path so that visual observations stay consistent with physics; commercial tiers can import 3D Gaussian-splatting scenes reconstructed from real video and assign collision meshes to them.

Motphys reports a 100K physics-steps-per-second figure on 1,024 parallel G1 humanoids on an AMD Ryzen 9 9950X, and million-steps-per-second-class throughput at 4,096 parallel environments in its own technical report. In a G1 whole-body tracking task held constant across observation, reward, algorithm (fastsac) and hyperparameters, the company reports an end-to-end wall-clock time of 10.2 minutes with MotrixSim versus 130.8 minutes for Holosoma on an Isaac Sim backend — a 12.8× figure. The company's broader marketing claim is a 3–10× training speedup; the two numbers come from different benchmarks and are recorded separately in DISCREPANCIES. Additional published stress tests include a LiDAR throughput of roughly 124 million rays per second on a 32-thread AMD path, a 100% grasp pass rate on a Franka arm, 92-link / 36-loop Strandbeest walking, and a 98.42% success rate with 0.5 mm error on a Robometrics motion-planning benchmark of 2,590 tasks. All such figures are company-reported and have not been independently reproduced.

MotrixLab is the imitation- and reinforcement-learning training framework, offering an open-source community edition for academic use and a commercial edition. MotrixGen is the data-synthesis pipeline: given a task description it generates expert trajectories at the hundred-thousand scale with 35-plus domain-randomization dimensions and per-item quality filtering. MotrixStudio is a scene editor integrating asset and scene libraries with an AI-agent assistant, while Forge is a Blender-based tool that lets agents auto-generate physical scenes. An asset toolchain covers MJCF import, convex decomposition, asset validation, and commercial conversions for 3DGS, point clouds, CAD and USD, plus AutoMJCF for procedurally assembled factory-floor scenes.

The engine and framework have been adapted to a broad compute ecosystem: NVIDIA, AMD and Apple abroad, and Kunpeng, Ascend, Hygon, Iluvatar, Moore Threads and Muxi domestically. The company says the engine's in-house core makes such ports comparatively fast. Real-world validation is claimed on Unitree Go1/G1 and ANYmal-C quadrupeds/humanoids, and the Wuhan training ground reports that Unitree, RealMan, Galaxea, Booster and Noetix devices have been connected for trial training.

Data Strategy and the "Virtual Training Ground"

Motphys's central thesis is that the bottleneck in embodied AI is data, and that simulation is the only way to clear it at acceptable cost. CEO Cui Hanqing has argued that real-world data collection is labor-intensive — it requires people, equipment, venues and time — whereas simulation is compute-intensive, generating data by consuming electricity and parallelism. He has publicly predicted that the data mix will shift toward 90–99% simulated data, with real data reduced to a small but indispensable residue that helps robots "see" real scenes while simulation teaches them to "learn." This is a company position, not an industry consensus; the report records it as such.

The practical mechanism is the Real2Sim2Real loop. In the "Real to Sim" direction, the company reconstructs real scenes 1:1, uses a physical-benchmark lab to measure real physical parameters, and offers tools to turn photos, video, point clouds, CAD and USD assets into simulatable bodies with collision and Physical attributes. In the "Sim" stage, the engine randomizes friction, mass, terrain and appearance across dozens of dimensions and runs thousands of parallel environments to accumulate interaction experience cheaply. In the "Sim to Real" stage, the goal is zero-shot transfer to the physical machine, avoiding costly on-robot tuning. The SimReady asset system is intended to make 3D assets carry mass, friction and joint constraints "at birth," so they are directly usable in training.

Cui has publicly argued that NVIDIA's USD-based SimReady definition is insufficient because it guarantees an asset can run in an engine but not that it improves model performance; Motphys says it is extending the concept toward measurable model benefit, while promising "weightier" results to come. That is a forward-looking claim without a published artifact at the time of writing. The report therefore treats the data flywheel as a coherent strategy whose economics are plausible but whose realized scale, customer retention and model-level gains remain unverified by independent evidence.

Benchmarking and Public Evaluations

Motphys has published an unusually detailed self-benchmark around the MotrixSim 1.0 launch at GLOSIM 2026 and COSCon 2026, framed as ten "devil-level" stress tests across architecture, contact fidelity, rendering, multimodal feedback, generation diversity and parallel scale. The headline results — sub-0.1% joint error, 100% Franka grasp pass rate with 0 mm slip at the friction threshold, stable Strandbeest walking where conventional engines diverge at a 5 ms step, 124 million LiDAR rays per second, and 98.42% planning success at 0.5 mm error — are presented as reaching an "industry ceiling."

The most consequential comparison is the G1 whole-body tracking task, where MotrixSim's 10.2-minute end-to-end time is benchmarked against Holosoma/Isaac Sim's 130.8 minutes under matched settings. This is the company's strongest quantitative argument for its efficiency claim, but it is a single task, a self-selected baseline and a company-run test. The report treats it as directional evidence of strong engineering rather than a settled, reproducible advantage. The academic record is more durable: a paper on GS-Playground, which fuses 3D Gaussian rendering with physics simulation, was accepted at RSS 2026, and the company says a second result was accepted at CoRL 2026. Peer-reviewed acceptance is a stronger signal than vendor benchmarks because it carries external review, though it does not by itself validate the product claims.

Business Model, Customers and Go-to-Market

Motphys serves three customer groups, per Cui: research institutions (for example the Shanghai national-local co-built humanoid robot innovation center), industry leaders (XCMG, leading automakers) and embodied-AI companies (D-Robotics, Booster). Commercialization runs through a free community edition that drives academic adoption and a commercial edition sold as licenses, custom bodies and scenes, add-on modules, on-premise deployment and engineering support with an SLA. The September 2026 joint solution with Huawei positions MotrixSim and MotrixLab as the simulation layer of Huawei's embodied-robot stack, with Huawei supplying compute and commercial reach — a channel strategy that reduces go-to-market cost but creates dependence on a single ecosystem partner that is also an investor.

The 2025 Wuhan training ground adds a project-based revenue line and a land-and-expand motion: the company and the district co-operate the facility, robot makers connect devices for trial training, and universities including Peking University and Harbin Institute of Technology have purchased the technology for teaching and research. Cui has framed the company as an "ecosystem connector" linking compute providers below and robot/OEM makers above. Revenue scale, contract values, customer count and the community-to-commercial conversion rate are not disclosed, which is the principal diligence gap in this report.

Competitive Landscape

Motphys competes at the simulation-infrastructure layer. Its most referenced rival is NVIDIA Isaac, which dominates through the PhysX engine, Omniverse and the USD-based SimReady standard, plus a deep, decade-long investment in simulation and a growing portfolio of acquisitions in the space. Google's MuJoCo is the other global reference point, especially for contact-rich research and as the format Motphys explicitly supports for asset import. Genesis, a newer physics simulator, and a cluster of world-model companies such as World Labs are reshaping the boundary between physically accurate simulation and learned world models; AMD's reported 2026 acquisition of World Labs and its acquisition of the robotics-simulator company SceniX signal that large platforms are converging on the same layer.

Within China, Motphys's differentiation claims are a fully proprietary engine core, broad domestic-chip adaptation and the lowest-level domestic control of the stack. It is not alone: other Chinese teams are building simulation or data platforms, and robot makers can in principle build their own simulators. Cui's own answer is that a high-quality simulation team is expensive and slow to maintain, so most companies will buy from specialists. The report judges the differentiation real but narrow: the moat lies in engine-level know-how and ecosystem positioning, and it will be tested as NVIDIA, Google and large platforms deepen their simulation offerings and as domestic peers mature.

Financing and Capital Structure

Motphys's investors are a notable signal for an infrastructure startup. Initial reporting in 2022 described an angel round of several million US dollars led by Sequoia China Seed and Hillhouse Ventures, with Game Trigger participating; later company materials list Sequoia China Seed, Hillhouse Ventures, MiraclePlus (Qiji Chuangtan) and Game Trigger. A 2024 strategic investment from Huawei's Hubble came alongside a business registration change adding Hubble-affiliated entities. In October 2026, one media report said a Pre-A round had completed closing with the amount undisclosed, while an interview published on 23 October 2026 said a new round was "about to close" — a conflict recorded in DISCREPANCIES. A third-party database tallies only two disclosed events totaling roughly RMB 71 million, a figure that clearly omits strategic and later rounds.

No reliable valuation for Motphys is public. This is material because several peers in the same CCTV "Winning AI+" cohort have public valuation figures in the billions of RMB, and investors in the space are pricing infrastructure companies heavily on narrative. For diligence, the items that would carry weight are a company statement on round structure and investors, signed customer contracts with values, and disclosed revenue or headcount trajectory; none is currently public beyond media summaries.

DISCREPANCIES

  1. Founding date and legal entity. Company website and some databases give 9 August 2021 for Beijing Mouxianfei Technology Co., Ltd.; 2022-era media and profiles describe a company founded in October 2020 under a similarly named entity. Ruling: likely an earlier operating entity versus a later re-registration; the 2020 origin is supported by multiple independent profiles, so both are retained and the exact legal-entity lineage is left open.
  2. Angel-round investor list. Early reports name Sequoia China Seed, Hillhouse Ventures and Game Trigger; later company material also lists MiraclePlus (Qiji Chuangtan), while a 2022 business-registration change did not show MiraclePlus or Game Trigger. Ruling: the four-institution list is the company's own current statement; Game Trigger and MiraclePlus are closely associated in Chinese reporting, so the discrepancy may be naming rather than substance — unverified.
  3. Financing stage and timing. A late-October 2026 media report says a Pre-A round completed closing (amount undisclosed); an interview published 23 October 2026 says a new round was "about to close." Ruling: these may describe different rounds or a closing-versus-announcement lag; the company has not published a formal round announcement, so the stage is recorded as unconfirmed.
  4. Disclosed financing total. A third-party database reports two events totaling roughly RMB 71.5 million, while media describe a several-million-US-dollar angel round plus a strategic investment and a Pre-A. Ruling: the database undercounts strategic and undisclosed rounds; it is not a reliable total.
  5. Training speedup: 3–10× versus 12.8×. Marketing text says 3–10×; the technical report cites a 12.8× end-to-end wall-clock advantage (10.2 vs 130.8 minutes) on one G1 task against Holosoma/Isaac Sim. Ruling: different benchmarks and baselines; neither is an independently reproduced, general-purpose speedup, so both are reported with their scope.
  6. Language about "domestic first / only." Company and partner materials call the platform the only domestic full-stack simulation chain, the first domestic virtual-physical integrated training center, and the first autonomous-controllable robot simulation-training ground. Ruling: these are first-party or partner claims about narrow categories; they are recorded as claims, not established market facts.
  7. Academic acceptances. Company materials say results were accepted at RSS 2026 and CoRL 2026, and one talk says "two top-conference papers this year." GS-Playground at RSS 2026 has an identifiable paper title; the CoRL result is described generically. Ruling: RSS acceptance is reasonably specific; the second paper's venue and title need verification before it is treated as established.
  8. Valuation. No authoritative figure is public; peers in the same cohort have public billion-RMB valuations. Ruling: valuation is left blank rather than inferred.

Risks and Due Diligence Gaps

Commercialization is the primary risk. Motphys has strong technical narratives and marquee investors but no disclosed revenue, contract values or community-to-commercial conversion data. The customer list is real but heavily weighted toward trial deployments and ecosystem partners, several of whom are also investors or government-linked programs. Dependence on a single large ecosystem partner — Huawei — for both compute adaptation and go-to-market creates concentration risk.

The technical claims, while impressive, are largely self-reported. Vendor benchmarks on self-selected tasks and baselines are not equivalent to independent reproducibility; a buyer should demand third-party evaluation. The "99% simulated data" thesis is a company view, and although simulation has clearly gained acceptance, the exact mix will depend on task domain and on evidence that simulated data improves model outcomes — the very gap Cui says Motphys is working to close. Competition from NVIDIA, Google and emerging world-model companies is intensifying, and the domestic engine talent pool is thin, which cuts both ways: high barriers to entry, but also a hiring and retention constraint.

Overall Assessment

Motphys is a technically credible, well-connected infrastructure company that has built a genuine full-stack simulation platform around a proprietary physics engine — a rare capability in China and a strategic asset as embodied AI scales. Its strengths are engine-level depth, a coherent Real2Sim2Real product story, broad domestic-chip adaptation, and an investor and channel position anchored by Sequoia, Hillhouse, MiraclePlus and Huawei. Its weaknesses are the absence of verifiable revenue, first-party-heavy performance claims, unresolved financing-stage reporting, and dependence on a single ecosystem partner. The report scores it 7.1 overall: a strong technology and product base whose commercial and financial case is still being proven.

References

[1] Motphys official site — product architecture, MotrixSim / MotrixLab / MotrixGen, asset toolchain (motphys.com). [2] Motphys "About Us" — founding, milestones 2021–2026, offices, compute ecosystem (motphys.com/zh/about/). [3] MotrixSim product page — physics fidelity, rendering, parallelism, community vs commercial editions (motphys.com/zh/motrixsim). [4] "MotrixSim 1.0 technical release" — GLOSIM 2026 / COSCon 2026 launch, benchmark figures, RSS 2026 / CoRL 2026 acceptances (ASIASIM / industry press, November 2026). [5] IPO Zaobaozhidao / Shangyexinzhi — interview with founder Cui Hanqing, financing history, SimReady and data views (23 October 2026). [6] CCTV Finance "Winning AI+" season 2 — 89.6 score, first place, founder remarks (9 October 2026). [7] RoboBaiku / WeChat — CCTV top-ten AI founders, financing and cohort analysis (30 October 2026). [8] CCTV Local / CRI — Wuhan Dongxihu simulation-training ground agreement, customer trial list, founder background (30–31 October 2025). [9] Yuanjie / WeChat — founder Cui Hanqing background, angel round, early company history (2022). [10] Yiou Data — company profile, financing events, headcount, registered address (data.iyiou.com). [11] Tsinghua MBA Embodied-AI Club forum summary — founder remarks on MotrixSim and RSS 2026, competitive views (May 2026). [12] China Daily / China Economic Net — company boilerplate, market deployment, certification and academic claims (2026).

This report is compiled from publicly available information, including the company's website, product documentation, technical reports, conference acceptances, partner case studies and media interviews. Figures cited under DISCREPANCIES are recorded separately by independent observation and are not merged. This report is for reference only and does not constitute investment advice. Data as of 15 November 2026.

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.