Fysics AI Series A

Fysics AI: The PhysX Founder's Second Act, a Domestic Differentiable Physics Engine and Its Ecosystem Play

A Shanghai physical-intelligence company founded by Zhang Lihua, a principal architect of the PhysX engine; it builds the Fysics differentiable engine, the OmniFysics foundation model and the MoziSim platform, closed an A1 round in August 2026, and released the unified world action model Fysiverse-UniWAM in late October 2026.

Executive Summary

Fysics AI, formally Fysics Intelligent Technology (Shanghai) Co., Ltd., is a physical intelligence company that landed in Shanghai's Xuhui district in 2025, founded by Professor Zhang Lihua of Fudan University. Zhang's record is the company's hardest technical credential: in 2003 at Ageia in the United States he led development of the first-generation PhysX physics engine and participated in designing the world's first physics acceleration chip, the PPU, and after NVIDIA acquired PhysX he led the engine's migration and optimization onto GPUs.

The product system has five parts: the differentiable physics engine Fysics, the embodied intelligence simulation and training platform MoziSim, the full-modality physical AI foundation model OmniFysics, the physical world model Fysiverse and the end-to-end physical AI data solution FysiData. On March 27, 2026 the company held a launch event in Shanghai titled from simulation to reality, opening a new era of physical AI, releasing Fysics and a full-stack technical foundation covering simulation platform, foundation model and evaluation benchmark.

Funding has been rapid. January 2026 brought a nearly 100 million renminbi Pre-A1 co-led by Matrix Partners China and Oriental Fortune Capital with Muxi and Chixing Ventures. February 2026 brought Pre-A2 led by Dingfeng Venture with Silgen Capital, Taiyou Fund and Yichang Capital, with existing investors adding. May 2026 completed Pre-A3, closing the Pre-A stage. August 2026 brought a Series A1 in the hundreds of millions of renminbi co-led by TEDA Capital, Haitian Capital and Fudan Venture with new investors including Yida Capital, Youshi Capital, Yongyue Capital, Huirong Fund, Caixin Capital and Xiangjiang Guotou, and existing investors including Yunqi Partners, Silgen Capital, Zhongying Venture and ChangLei adding.

Technical and ecosystem progress includes FysicsWorld in December 2025 with Fudan's Cognitive and Intelligent Technology Laboratory, described as the world's first unified full-modality benchmark for the real physical world; OmniFysics in February 2026, described as the world's first full-modality physical AI foundation model that understands physical laws at only three billion parameters; Fysiverse in June 2026 moving from visual generation to physical inference; and at WAIC 2026 the formation of China's first physical intelligence innovation consortium with 32 universities, companies and investors including Fudan, Harbin Institute of Technology, Muxi, Topstar, BRL and the China Software Testing Center.

Open-source and release activity accelerated in the second half of 2026. On September 24 the company open-sourced three things at once: Fysiverse-3D for building executable 3D worlds, the OmniFysics twins consisting of the OmniFysics-Captioner video description model and the 4B-parameter OmniFysics-Nano-V2, which the company says took first place on 24 international benchmarks together with 7 for Captioner and 17 for Nano-V2, and FysicsReason, described by the company as the first benchmark for full-modality world perception and physical state reasoning, splitting perception from reasoning through world state variables. From October 31 to November 1, 2026 the company and Fudan's CITLab released Fysiverse-UniWAM, a unified world action model placing navigation, mobile manipulation and fixed-station manipulation in one model, with the architecture, training and inference code, navigation weights and dataset released.

In physical AI terms Fysics AI positions as infrastructure. Zhang states explicitly that the company is not a model company but an infrastructure company selling foundations others can build on, spanning engine, platform, standards and ecosystem. The underlying judgment is that competition in physical AI has shifted from parameter counts to coordinating a whole ecosystem, and that like early autonomous driving the competition will move from algorithms to data loops, engineering systems and infrastructure.

One framing point matters. This is a company with a top-tier founding record, extremely fast fundraising and a very broad product matrix. Its strengths are Zhang's PhysX background in the world's most widely used physics engine, a differentiable engine route that differs from mainstream engines, and dense industrial partnerships with Topstar, BRL and Muxi. Its weaknesses are absent product validation across five early products, undisclosed revenue and the verifiability of repeated first-in-the-world claims.

Company Overview

Team and Founder

Zhang's record is the core asset. Leading PhysX development at Ageia in 2003 and participating in the PPU design, then leading GPU migration after NVIDIA's acquisition, gave him two things: deep understanding of engine internals such as collision detection, rigid body solving and constraint solving, and judgment about how engines evolve with hardware.

He is now a professor at Fudan and published a signed article in People's Daily in May 2025 asking why robots should be more human. Core members come from Tsinghua, Peking University, Fudan and Shanghai Jiao Tong, covering physics simulation, embodied intelligence and multimodal models.

From Municipal Project to Xuhui Chain Leader

In July 2025 the company signed as a major Shanghai project at the World Artificial Intelligence Conference and settled in Xuhui, where it is now chain leader of the physical intelligence industry chain. That role means not only receiving policy support but acting as industry organizer through consortiums, standards and scenario access.

In April 2025 an embodied intelligence simulation platform built on Fysics started. At WAIC 2026 the company led formation of the Physical Intelligence Innovation Consortium with 32 members.

Technology

What a Differentiable Physics Engine Means

Differentiable physics is the central technical claim and deserves explanation.

Conventional engines such as PhysX, Bullet and MuJoCo compute forward: given state and forces, produce the next state. That suffices for simulation but not optimization, since adjusting parameters to make a robot complete a task requires knowing how parameter changes affect outcomes, which is the gradient.

A differentiable engine makes the whole pipeline differentiable so that control, material and structural parameters can be optimized by gradient descent. The value in robotics is concrete: policy learning can use gradients rather than sampling with orders of magnitude better efficiency, and domain randomization for sim-to-real can search parameter space with gradients.

Three difficulties arise. Contact and collision are non-smooth, with forces jumping between contact and separation so gradients are discontinuous and require relaxation or smoothing. Long horizons cause gradient explosion or vanishing through rollout. And performance overhead grows because intermediate states must be stored for backpropagation.

The company calls Fysics China's first fully differentiable physics engine, a first whose verifiability is limited by how fully differentiable is defined, though the direction matches international frontiers such as NVIDIA Warp, DiffTaichi and Brax.

OmniFysics at Three Billion Parameters

OmniFysics is described as the world's first full-modality physical AI foundation model understanding physical laws at only three billion parameters. Two terms need unpacking.

Full-modality means handling vision, language, sensors and actions with reasoning in a shared representation. Understanding physical laws means predictions obey constraints such as conservation, causality and contact rules rather than merely fitting visual data.

Three billion parameters is a deliberate choice. Against language models of hundreds of billions to trillions, three billion is small, but a physics model needs efficient representation of physical laws rather than language knowledge, and that scale suits edge or workstation deployment on robots and training stations.

Understanding physical laws is a strong claim. Constraints can enter through loss functions as soft constraints, network structure as hard constraints such as equivariant networks, or post-processing correction, and the three differ greatly in effect without disclosure of which is used.

Fysiverse and Physical Inference

Fysiverse released in June 2026 is the physical world model, described as moving from visual generation to physical inference.

The distinction matters: visual generation learns what looks right while physical inference learns what will happen. The former can produce convincing video that is physically wrong, with objects interpenetrating, gravity misdirected or collisions non-conservative, while the latter requires predictions consistent with physics.

The direction matches World Labs and AMI Labs, with the difference that Fysics AI starts from an engine guaranteeing physical correctness and adds generation, while World Labs starts from generation and adds physics. Engine-first gives stronger physical correctness with potentially weaker scene diversity and visual quality, and generation-first reverses the trade-off.

Fysiverse-UniWAM as a Unified World Action Model

From October 31 to November 1, 2026 the company and Fudan's CITLab released Fysiverse-UniWAM, whose goal is to cover where to walk, where to stop and what to do next inside a single model.

Two design choices carry most of the weight. The first is MAP (Manipulation Anchor Pose), which describes the pose the chassis should reach before manipulation starts in terms of a two-dimensional position and heading on the ground, backed by an automated pipeline that derives reference docking poses and approach trajectories from 3D scenes, target positions and traversable regions, with MAP-Data holding more than 1.5 million trajectories and about 7,500 hours. The second is the Mixed-Stream two-stream architecture, where the navigation stream and the manipulation stream keep their own visual inputs, action encoders and output heads while sharing video and action backbones, and where the two sample types are drawn independently and trained jointly, which avoids both the interface mismatch of forcing two action types into one action vector and the handover error of training two separate policies.

The reported metrics need to be read with their conditions. On MAP-Bench the UniWAM-Scaled variant reaches a mean position error of 9.85 cm and a mean heading error of 2.30 degrees, which the company reports as 30.1 percent and 44.0 percent below the strongest external baseline. On real robots the evaluation spans four platforms and 24 tasks: all 13 MAP navigation tasks rank first on position error, the four mobile manipulation tasks rank first on success rate with an average of 80.0 percent, and the seven fixed-station manipulation tasks reach an overall success rate of 68.9 percent, 2.5 percentage points behind the baseline. A separate zero-shot figure comes from UniWAM-Sim, trained only on MAP-Data with no real robot data, which reaches 18.7 cm position error and 6.34 degrees heading error on hardware, clearly above the 9.85 cm benchmark figure because the data sources and test conditions differ, so the two pairs should not be merged into one error level. Released artifacts include the Mixed-Stream architecture, training and inference code, the UniWAM-Base and Scaled navigation weights, and the MAP-Bench data and evaluation trajectories, with the technical report at arXiv 2609.39388.

FysicsWorld as a Benchmark

FysicsWorld released in December 2025 with Fudan's laboratory is described as the world's first unified full-modality benchmark for the real physical world.

Benchmarks matter more than often recognized, since a field's maturity depends on accepted evaluation allowing methods to be compared and progress measured. Physical AI lacks accepted benchmarks with each company using its own test sets, and if adopted FysicsWorld becomes an asset conferring standard-setting voice.

Adoption requires time and community consensus, and whether a company-led benchmark becomes an industry standard depends on openness, including open sourcing, multi-party submission and transparent evaluation.

Technical Shortfalls

First, maturity across five products, all early with almost no public technical metrics on solver accuracy, speed or model performance.

Second, engineering difficulties in differentiable simulation, where non-smooth contact, long-horizon gradients and memory overhead have academic solutions but unverified stability at industrial scale with many objects, materials and large scenes.

Third, the capability boundary of a three-billion-parameter model for generalization to new scenes, materials and tasks.

Fourth, verifiability of the repeated first claims without third-party confirmation.

Product and Commercial Model

Product Division

Fysics is the engine layer, MoziSim the platform for embodied simulation and training, OmniFysics the model layer, Fysiverse the world model and FysiData the data layer, forming a coherent chain from engine through platform, model, world model and data.

From September 24, 2026 the company added open-sourced modules to that chain: Fysiverse-3D for building executable 3D worlds, made up of the verifiable spatial text-to-image scheme SpatialGuard, the unified spatial reasoning model Fysiverse-3D-Vision and the simulation-ready framework Fysiverse-3D-SimReady; the OmniFysics twins, where Captioner describes physical processes and Nano-V2 reasons about physical laws; and the FysicsReason evaluation benchmark.

Fysiverse-UniWAM, released with Fudan's CITLab from October 31 to November 1, 2026, adds an execution layer outside that stack: it unifies navigation, mobile manipulation and fixed-station manipulation in one world action model built on the MAP design and the Mixed-Stream two-stream architecture.

The product count needs a stated basis. The rest of this report still refers to the five products presented at the March 2026 launch event, while the two releases since September 24 leave more than five separately countable modules, so the two figures describe different points in time rather than one inventory.

Commercialization Path

Four target tracks are industrial robotics, humanoid embodied intelligence, unmanned mining machinery and autonomous driving simulation, all sharing intensive physical interaction, expensive real data collection and high simulation value.

The three uses of Series A1 proceeds confirm the path: deep iteration of engine solving performance across rigid, soft, fluid and granular multiphase coupling; product and industry solution delivery including MoziSim enterprise commercialization and standardized solutions in four tracks; and ecosystem and team expansion.

Six Industrial Partnerships

Six partnerships announced at WAIC show concrete commercial forms: an industrial production line physical digital twin with Topstar, a zero-carbon mine high-risk scenario simulation system with BRL, a fully domestic FysiStation physical AI training workstation with Muxi, a simulation test field with a testing institution, a data collection closed loop and an industry-academia laboratory.

Their natures differ, with Topstar and BRL as industry customers, Muxi as compute partner for domestic adaptation, the testing institution as standards partner and the laboratory for talent, showing the company working simultaneously on technology, scenarios, compute, standards and talent.

Business Model and Pricing

The model is likely a mix of engine licensing, platform subscription and project delivery, licensing to robot makers and simulation vendors, subscription to R and D teams and projects to industry customers.

Pricing is undisclosed. Engine licensing has precedents, with PhysX once a free SDK plus hardware acceleration and Havok commercially licensed, and differentiable engines may price by call volume or seat.

Competitive Landscape

Physics Engine Rivals

Traditional engines include NVIDIA PhysX for games and robotics, MuJoCo as the de facto standard for robot reinforcement learning, Bullet as open source and Havok commercially. Differentiable engines include NVIDIA Warp, DiffTaichi, Brax from Google and the open-source Genesis in 2024.

Fysics AI differentiates on domestic substitution plus full stack, since a self-controlled differentiable engine carries policy value and the full stack offers one-stop convenience.

World Model Rivals

Fysiverse faces World Labs with Marble and Atlas, AMI Labs with JEPA and corporate world model research. Its differentiation is physics-first from an engine, with disadvantages in team size and capital, since World Labs was acquired by AMD for 8.2 billion dollars and AMI Labs raised a 1.03 billion dollar seed.

Embodied Simulation Platform Rivals

MoziSim faces NVIDIA Isaac Sim with the most mature ecosystem, MuJoCo, Genesis and domestic platforms. Isaac Sim leads on ecosystem including asset libraries, sensor models and Omniverse integration, while MoziSim differentiates on domestic substitution, a differentiable engine and adaptation to domestic compute such as Muxi.

Benchmark Influence

If adopted, FysicsWorld becomes soft power, though benchmark competition is intense with many self-published benchmarks, and the winner is usually the most open and trusted by the community.

Financing and Capital Structure

Funding Pace

Four rounds in 2026 from Pre-A1 at nearly 100 million renminbi through Pre-A2, Pre-A3 and a Series A1 in the hundreds of millions indicate both strong investor enthusiasm, especially from industrial and local funds, and heavy capital needs for team, product and ecosystem expansion.

The investor mix combines market venture capital with industrial capital, local and university funds and specialist funds, with Fudan Venture's participation tied to Zhang's Fudan affiliation.

Valuation and Use of Proceeds

Valuation is undisclosed. With a Series A1 in the hundreds of millions and four rounds in a year, cumulative funding is likely in the several hundred million to one billion renminbi range with valuation plausibly in the billions.

Proceeds go to technology iteration, product and industry delivery and ecosystem and team expansion.

DISCREPANCIES

1. Founding versus landing dates. Landing in Shanghai's Xuhui district was in July 2025 at WAIC, while registration may be earlier, and some material states founding in 2025. This report uses landing in Shanghai in 2025 without inferring registration. Source tier: OFFICIAL (signing fact) with registration information absent.

2. Density of funding rounds. Four rounds in one year from Pre-A1 through A1 are sometimes simplified to one A round. This report lists all four per the official milestone page. Source tier: OFFICIAL (company milestones).

3. The Series A1 amount in the hundreds of millions. The company says hundreds of millions of renminbi without a precise figure, while some media write nearly 100 million or tens of millions of dollars. This report uses the hundreds of millions framing and notes imprecision. Source tier: OFFICIAL (company statement).

4. Dense use of first-in-the-world and first-in-China claims. The company applies first claims to OmniFysics, Fysics, FysicsWorld and the consortium, all depending on definitional boundaries and not independently verifiable. This report retains them labeled as company framing. Source tier: company marketing, UNVERIFIED.

5. OmniFysics parameters versus capability. The company says three billion parameters achieve innovative breakthroughs in physical perception and reasoning without disclosing benchmark results. This report labels parameters and positioning disclosed with performance absent. Source tier: OFFICIAL (parameters) with performance absent.

6. Definition of fully differentiable. Fully differentiable may mean the entire pipeline is differentiable or all physics types support differentiation, which differ in substance. This report adopts the pipeline reading. Source tier: OFFICIAL (statement) plus inference.

7. The PhysX founder framing. The company calls Zhang a core founder of NVIDIA's PhysX engine, while strictly PhysX originated at Ageia in 2003 and NVIDIA acquired it in 2008, with Zhang leading development in the Ageia period. This report uses the accurate Ageia-period formulation. Source tier: public biography (MEDIA) correcting company framing.

8. Composition of the 32-member consortium. The company says the consortium brings together 32 universities, enterprises and investors including Fudan, Harbin Institute of Technology, Muxi, Topstar, BRL and the China Software Testing Center, without publishing the full list or the depth of participation such as funding or joint projects. This report labels it as led formation without inferring commitments. Source tier: OFFICIAL (formation fact).

9. Timing and framing of the open-source release and UniWAM. The three open-sourced pieces (Fysiverse-3D, the OmniFysics twins and FysicsReason) were released on September 24, 2026, while Fysiverse-UniWAM came on October 31 and November 1, 2026, and some coverage merges the two into a single open-source and launch event. This report records the two dates separately. Source tier: official technical report crossed with media coverage (OFFICIAL and MEDIA).

10. Two sets of error figures for UniWAM. The 9.85 cm mean position error and 2.30 degree mean heading error are UniWAM-Scaled averages on the MAP-Bench evaluation set, whereas the 18.7 cm and 6.34 degree figures come from UniWAM-Sim trained only on MAP-Data with no real robot data and deployed zero-shot on hardware. Data sources and test conditions differ, so the two pairs should not be merged into a single error level. Source tier: technical report tables (company-published, not independently reproduced).

11. Product count. Several passages still say five products, which reflects the product matrix at the March 2026 launch event, while after the September 24 open-source release and the October 31 to November 1 UniWAM launch more than five modules can be counted separately. The two numbers refer to different points in time, so this report keeps the historical five-product wording and records the difference here. Source tier: company milestones and release materials (OFFICIAL).

Risks and Diligence Gaps

First, absent product validation. All five products are early with no public accuracy, speed or performance metrics, while the technical value of a differentiable engine and a physical foundation model needs data.

Second, sustainability of four rounds in a year. Fast fundraising shows enthusiasm and also fast burn, and disappointing commercialization would make the next round difficult.

Third, an over-broad portfolio. Engine, platform, model, world model and data advancing simultaneously strains a young company.

Fourth, engineering stability of differentiable simulation, with non-smooth contact, long-horizon gradients and memory overhead unverified at industrial scale.

Fifth, ecosystem dependence on Muxi for domestic compute, Topstar and BRL for scenarios and Fudan for research.

Sixth, competition, with NVIDIA holding strong positions in PhysX, Warp and Isaac and domestic peers moving quickly.

Seventh, the trust cost of repeated first claims, which erodes credibility especially without technical metrics.

Composite Assessment

Technology scores 7.5. Zhang's PhysX record is a real asset, the differentiable engine route matches international frontiers and is scarce domestically, the five-layer architecture from engine to data is coherent, and FysicsWorld has community value. Deductions for undisclosed metrics, unverified engineering stability and an unclear capability boundary at three billion parameters.

Business scores 6.5. Four target tracks have clear demand, six industrial partnerships provide landing paths and the Muxi partnership solves domestic compute adaptation. Deductions for undisclosed revenue, early products and unvalidated commercialization.

Finance scores 7.0. Four rounds in a year from Pre-A1 at nearly 100 million to a Series A1 in the hundreds of millions, with market venture, industrial, local and university investors, provides ample capital. Deductions for fast burn, undisclosed valuation and uncertain sustainability.

Product scores 7.0. Five products cover the physical AI stack, MoziSim enterprise and four-track solutions are progressing and FysiStation solves domestic compute deployment. Deductions for low maturity, undisclosed interface and usability and cases without quantification.

Market scores 7.5. Physical AI draws policy and capital attention, embodied intelligence and smart manufacturing demand are real and the Xuhui chain leader role provides scenarios and policy resources. Deductions for intense competition with a strong NVIDIA ecosystem, high market education cost and long adoption cycles.

Overall 7.1. This is the domestic physical AI company with the hardest founding record, the broadest layout and the least validation: Zhang's PhysX background provides technical credibility, the differentiable engine plus physical foundation model plus world model provides narrative completeness and four rounds in a year provide resources. Yet all five products are early and metrics and customer cases are missing. The judgment on physical AI is that the technical direction of differentiable simulation plus physical foundation models is correct and at the international frontier, while moving from product launch to industrial usability requires clearing engineering stability, performance validation and a commercial loop.

References

  1. Fysics AI milestones from the company website: April 2025 embodied intelligence simulation platform launch; May 2025 chairman Zhang Lihua's signed People's Daily article; July 2025 signing as a major Shanghai project at WAIC and settling in Xuhui; December 2025 FysicsWorld benchmark with Fudan's laboratory; January 2026 nearly 100 million renminbi Pre-A1 co-led by Matrix Partners China and Oriental Fortune Capital with Muxi and Chixing Ventures; February 2026 Pre-A2 led by Dingfeng Venture and release of OmniFysics; March 27, 2026 launch of Fysics and the full-stack foundation; May 2026 Pre-A3; June 2026 Fysiverse; August 2026 Series A1
  2. Series A1 announcement from CNR, China Business Journal and China Internet Information Center, August 6, 2026: hundreds of millions of renminbi co-led by TEDA Capital, Haitian Capital and Fudan Venture with new investors Yida Capital, Youshi Capital, Yongyue Capital, Huirong Fund, Caixin Capital and Xiangjiang Guotou and existing investors Yunqi Partners, Silgen Capital, Zhongying Venture and ChangLei adding; prior investors including Matrix Partners China, Oriental Fortune Capital, Yunqi Partners, Dingfeng Venture and Zhidao Capital plus Muxi, Chixing Ventures, Taiyou Fund, Yichang Capital, Changshi Capital and Pangu Venture
  3. Five core products from the Series A1 announcement: differentiable physics engine Fysics, embodied intelligence simulation and training platform MoziSim, full-modality physical AI foundation model OmniFysics, physical world model Fysiverse and end-to-end data solution FysiData, described as China's first fully self-developed full-stack physical intelligence technology system
  4. Team and founder from public reporting: Zhang Lihua leading development of the first-generation PhysX engine at Ageia in 2003 and participating in the PPU design, later leading GPU migration after NVIDIA's acquisition, now a Fudan professor and company chairman; core members from Tsinghua, Peking University, Fudan and Shanghai Jiao Tong
  5. Consortium and partnerships from the Series A1 announcement and media: China's first Physical Intelligence Innovation Consortium formed at WAIC 2026 with 32 universities, enterprises and investors including Fudan, Harbin Institute of Technology, Muxi, Topstar, BRL and the China Software Testing Center; six partnerships covering an industrial production line physical digital twin with Topstar, a zero-carbon mine high-risk scenario simulation system with BRL, a fully domestic FysiStation training workstation with Muxi, a simulation test field, a data collection closed loop and an industry-academia laboratory
  6. Use of proceeds from the Series A1 announcement: deep iteration of core technology including Fysics solving performance across rigid, soft, fluid and granular multiphase coupling and iteration of OmniFysics and Fysiverse; product and industry delivery including MoziSim enterprise commercialization across industrial robotics, humanoid embodied intelligence, unmanned mining machinery and autonomous driving simulation; ecosystem and team expansion
  7. Positioning from public reporting: Zhang describing the company as an infrastructure company rather than a model company, selling foundations others can build on across engine, platform, standards and ecosystem, and arguing that physical AI competition has shifted from parameter counts to coordinating a whole ecosystem
  8. Technical judgment from public reporting: Zhang comparing physical AI to early autonomous driving where competition moves from algorithms to data loops, engineering systems and infrastructure, requiring quantifiable, reproducible and continuously verifiable engineering systems linking simulation, training, validation and deployment
  9. OmniFysics release from the company milestone page, February 2026: described as the world's first full-modality physical AI foundation model that understands physical laws at only three billion parameters, targeting embodied intelligence and humanoid robots
  10. Technical background from public sources: differentiable physics simulation makes the pipeline differentiable so control and material parameters can be optimized by gradient; difficulties include non-smooth contact and collision with discontinuous gradients, gradient explosion or vanishing over long horizons and memory overhead from storing intermediate states; international work includes NVIDIA Warp, DiffTaichi, Brax and Genesis
  11. Competitor information from public sources: traditional engines include NVIDIA PhysX, MuJoCo, Bullet and Havok; world model competitors include World Labs with Marble and Atlas acquired by AMD for about 8.2 billion dollars in 2026 and AMI Labs with JEPA and a 1.03 billion dollar seed; embodied simulation platforms include NVIDIA Isaac Sim, MuJoCo and Genesis
  12. Benchmark background from public sources: physical AI lacks accepted evaluation standards with each company using its own test sets; adoption depends on openness including open sourcing, multi-party submission and transparent evaluation, and a company-led benchmark needs community consensus to become an industry standard
  1. Company technical report page at physics.net/report.html and the UniWAM technical report at arXiv 2609.39388: the September 24, 2026 open-sourcing of Fysiverse-3D, the OmniFysics twins and FysicsReason, and the October 31 to November 1, 2026 Fysiverse-UniWAM report with MAP-Data and MAP-Bench metrics
  2. Jiemian, November 1, 2026, on the release of the unified world action model Fysiverse-UniWAM: the MAP and Mixed-Stream two-stream designs, more than 1.5 million trajectories and about 7,500 hours of data, the composition of the 24 real-robot tasks and the list of released artifacts
  3. Synced, November 1, 2026, on Fysiverse-UniWAM: the 9.85 cm and 2.30 degree MAP-Bench figures, the 80.0 percent average mobile manipulation success rate, the 68.9 percent fixed-station success rate, and the 18.7 cm and 6.34 degree zero-shot transfer results
  4. Jiemian, September 25, 2026, and the Fudan Lin-Gang industrialization platform account: the September 24, 2026 open-sourcing of Fysiverse-3D, OmniFysics-Captioner, OmniFysics-Nano-V2 and FysicsReason, with 7 first places for Captioner and 17 for Nano-V2
This report is compiled from public sources for informational purposes only and is not investment advice. The company is private and revenue and product performance are undisclosed; multiple first-in-the-world and first-in-China claims are company framing without third-party verification, and related judgments are labeled as uncertain. Data is current as of 2 November 2026.

Related reports66

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.