Legal name: Hangzhou Zhuoyin Intelligence Technology Co., Ltd.
JoinAI Deep Dive: From Synthetic Training Data to the Experience Layer for Physical AI — Rebuilding the Starting Point of Machine Learning with In-House World Models
JoinAI is a Hangzhou-based Physical AI startup building an 'Experience Layer' with its in-house diffusion model Simulaix and world model Terra. It serves more than a dozen clients including OpenAI and Google, and has closed an angel plus round backed by two state-linked funds.
Executive Summary
JoinAI (the brand of Hangzhou Zhuoyin Intelligence Technology Co., Ltd.) is a Chinese Physical AI startup that builds training infrastructure around generative world models. Founded by Kaiwen Yu, a PhD who previously served as a senior AI architect at Ant Group and Baidu, the company started in synthetic training data for autonomous driving and has since repositioned itself as an "Experience Layer" for Physical AI — a supplier of task-level experience that machines need in order to learn real-world tasks, rather than a supplier of images or video alone.
The product stack has three layers. Terra is the in-house generative world model: Terra-1 (2023) is a Transformer-plus-diffusion world model aimed at physically plausible, interactive driving scenes, while the newer Terra-3 adds spatial composition and object-level editing, turning a text prompt into an editable spatial scene with attached camera parameters and action annotations. Simulaix (2022) is the diffusion-based data-production platform that generates images together with pixel-level annotations and compresses a seven-step data pipeline into three. Chain is the data-management layer that unifies real and generated data, labels, versions and model feedback. Company materials now frame these as one closed loop that runs from identifying a model's capability gap to producing, scaling and verifying the experience that closes it.
JoinAI reports commercial validation with more than a dozen clients including OpenAI, Google, SimplifyNext, Zhongke Chuangda and Zhonghao Xinying, and says its technology covers more than 20 commercial deployment scenarios across more than ten countries. Its disclosed POC results are unusually concrete for a synthetic-data vendor: in L4 urban segmentation, mixing 10K generated images with 10K road-collected images raised mIoU from 45.7 to 48.9; in a long-tail cone scenario, 5K generated images alone reached 78.8 mIoU versus 36.75 for 0.5K real images. These numbers are company-published and customer-authorized, but not independently reproduced.
This report treats JoinAI as an early-stage infrastructure company with credible engineering and a coherent strategic narrative, whose decisive open questions are whether "effect delivery" and Learning Gain verification can be repeated across customers, and whether its world-model route can hold a position between traditional simulation vendors and large-platform world foundations. Financing comprises a million-yuan-scale angel round led by Angel Bay Capital in 2024 and a subsequent tens-of-millions angel-plus round backed by Hangzhou Talent Fund and Hangzhou Runmiao Fund. No authoritative valuation has been disclosed.
Company Overview
Public records and company materials place the operating entity's registration on 1 December 2021 in Hangzhou's Binjiang District, with the Zhuoyin Intelligence brand established in 2022 and a separate financing vehicle, Hangzhou Zhuoyin Zhixin Technology Co., Ltd., appearing from 2024. This multiplicity of entities is a recurring source of conflicting dates across databases and media, and is handled item by item in the Discrepancies section.
Founder, CEO and CTO Kaiwen Yu holds a doctorate from Newcastle University and is described by investors as having worked in AI since 2010, with senior architect roles at Baidu and Ant Group. Company materials list a compact leadership team: Xiaomeng Jiang (COO and product), Yang Zhou (CMO and spatial intelligence VP), Run Wu (engineering VP), Shengfan Ke (project VP), Xiaogang Xiang (algorithm VP) and designer Chengfang Yang, with core members drawn from Alibaba, Baidu, NetEase and Google. The company presents itself as a team of engineers with more than a decade each in AI.
Geographically the company is Hangzhou-based and has been incubated through funds and incubators linked to Hangzhou Capital, including a Xiaomi-ecosystem AIoT innovation center. Its stated mission — "better data, better AI, for everyone, forever" — and its habit of publishing long, first-person founder essays give it a distinctly narrative-driven public presence. That style is useful for understanding intent but must be separated from verifiable delivery when assessing the company.
Technology
JoinAI's technical route is explicitly generative rather than simulator-centric. Company materials argue that traditional simulation first defines a world with manual scenes, rules and parameters and then runs it, so the simulator remains the organizing center; adding generative AI to that framework only improves asset and scene construction without changing the underlying bottleneck. JoinAI instead proposes to start from real business tasks, identify the experience a model lacks, and generate the scenes and signals that supply it.
Terra-1 is described as a world model specifically designed to simulate reality, with emphasis on physically plausible, interactive and highly realistic autonomous-driving environments. Its architecture accepts multiple modalities — video, audio and images — encodes them separately, unifies them into tokens, and feeds a Transformer-plus-diffusion world model that predicts the next frame to produce continuous video. Company descriptions say action inputs can influence outcomes and that the model covers multiple sensor modalities including RGB and infrared, with control over variables such as ethnicity, pose, working conditions and environment.
The newer Terra-3 shifts emphasis from driving scenes to spatial composition and object-level editing. Marketing shows a single prompt producing an editable spatial scene, with "add a glass and place it on a magazine" as the illustrative edit. More importantly for training, the same learnable world can output several synchronized data types: image data, video data, camera parameters in structured JSON (extrinsics, intrinsics, sensor width and focal length) and action annotations that break a task into timed steps such as pick, transport and place for a robotic pick-and-place clip. This multi-output design is the technical substance behind the Experience Layer claim: one scene yields pixels, geometry, camera geometry and temporal action structure together.
Simulaix, introduced in 2022, is a self-developed diffusion model for label-to-training-data generation. Its defining trick is to extract pixel-level object masks and region information during image generation rather than in a separate annotation pass. The company claims this compresses a conventional seven-step data pipeline — scene preparation, collection, transfer and storage, preprocessing, annotation and so on — into three steps of prompt input, generation and data filtering, reducing acquisition cost by 80 to 90 percent and, in earlier materials, delivering data at under one tenth of the cost and more than ten times the speed. Simulaix supports batch generation, multi-sensor simulation beyond 2D cameras, and custom controls over label types, data appearance and the generation simulator.
Chain is the management layer that unifies real and generated data, labels, versions and model feedback into a closed asset loop. Together the three products map to a division of labor the founder describes plainly: Terra generates, Simulaix produces, Chain manages and feeds back. The company also ships standardized off-the-shelf datasets — urban road driving, vehicle fisheye parking and industrial hot-work safety among them — and a data platform at dataset.join-intelligence.com.
Product and Commercial Model
JoinAI's commercial narrative has evolved from delivering training-ready data toward what it calls effect delivery. The stated process runs from a business problem to a model capability gap, then to Experience design, data generation and collection, training, effect verification and the next iteration. The company is explicit that a customer cares less about how many frames or terabytes were delivered than about why a model failed, which experience it lacks, and whether adding new training data actually improved the model. It coins the term Learning Gain for this verified improvement and says it is moving from data delivery into the customer's training loop.
Disclosed POC results, published with customer authorization, are the strongest proof points. In L4 urban-road semantic segmentation over twenty object classes, 10K road-collected images scored 0.874 precision and 45.7 mIoU; 10K generated images scored 0.844 and 41.8; the 20K mix scored 0.92 and 48.9. In a long-tail cone scenario, 0.5K real cone images reached 0.31 precision and 36.75 mIoU, 5K generated cone images reached 0.94 and 78.8, and the mix reached 0.95 and 81.4. In an L2.5 fisheye pedestrian-detection test with YOLOv8x, 5K real images reached 0.84 precision and 0.82 recall, 5K generated reached 0.86 and 0.84, and the mix reached 0.91 and 0.89. Investor materials summarize this as a rise from 0.87 to 0.92 when generated and real data are mixed, and a database entry claims Simulaix improved client model performance by more than ten percent by late 2023. These figures are internally consistent and directionally plausible, but all originate with the company and its customers; no third-party benchmark, test protocol or customer contract has been published.
The customer list is the other commercial anchor. Company materials name OpenAI, Google, SimplifyNext, Zhongke Chuangda, Zhonghao Xinying and several Chinese partners, and say business spans autonomous driving, intelligent cockpits, industrial and special scenarios, robot manipulation, dexterous industrial tasks, service robots and world models. Investor accounts add Huawei, Didi, Xiaomi, Ant Group, Ecovacs and Desay SV. The company says it supports paid POC cooperation, sometimes starting verification at a first meeting, and that it had won tens of millions of yuan in orders from leading autonomous-driving and robotics clients by May 2024. Revenue, contract values and the conversion rate from POC to long-term Experience Programs are not disclosed.
Competitive Landscape
JoinAI sits in a crowded and fast-moving layer between synthetic-data vendors and world-model developers. Its own materials position the competition in three groups. Traditional simulators are strong at controllable scene construction but suffer from long scene-building cycles, limited realism and a sim-to-real gap; generative enhancement of a simulator improves efficiency without changing its simulator-centric core. General AIGC image and video tools optimize for visual naturalness, while training demands controllability, accurate labels, reasonable data distribution, multimodal consistency and stable batch delivery. World-model and driving-world approaches — NVIDIA Cosmos, Waymo's world model, DeepMind Genie 3 and World Labs Marble — aim to generate how a world evolves under actions, which is adjacent to but broader than JoinAI's task-specific experience production.
JoinAI's differentiation is the combination of a fully in-house generative stack, a label-to-data pipeline and a management loop tied to customer training feedback, plus a claim of real-world anchoring: experience must originate in genuine task failures and capability gaps. This is a credible wedge in a market where generic synthetic images are commoditizing. The risk is that the same large platforms investing in world foundations can absorb scene generation and annotation as features, and that dedicated synthetic-data vendors and in-house robotics data teams compete on cost and domain depth. The company's response is that its value is not a fixed data format but the ability to define, produce, scale and verify the experience a task requires — a thesis that is strategically coherent but not yet backed by a durable, published moat.
Financing
JoinAI raised a million-yuan-scale angel round on 30 May 2024, reported as exclusively backed by Angel Bay Capital (Tisiwi). A later angel-plus round of tens of millions of yuan was backed by Hangzhou Talent Fund (Phase II) and Hangzhou Runmiao Fund, both linked to Hangzhou Capital, with the stated use of funds going to generative world-model R&D, Experience infrastructure, training loops, product engineering and overseas expansion. Credited sources variously label this later round "angel plus" and "Pre-A", and give dates ranging from July to September 2026, a discrepancy handled below.
The cap table shows meaningful founder control. Public registries describe Yu indirectly holding roughly 73.6 percent of the operating company through a chain of holding entities, and an earlier snapshot listed him at 99 percent before the 2024 financing and holding-company restructuring. No authoritative valuation has been published for JoinAI. For due diligence, the material documents would be the company's formal round announcement, signed client contracts with amounts, and revenue or headcount trends; at present these are available only through media aggregation and company narrative.
DISCREPANCIES
1. Founding date. Registries give 1 December 2021 for the operating entity; company brand materials and most media use 2022; one financing-vehicle profile says the entity was founded in 2024. Ruling: these refer to different legal entities — operating company, brand launch and financing vehicle — and are recorded separately rather than merged into one date.
2. Company name and legal entities. Materials alternate between Hangzhou Zhuoyin Intelligence Technology Co., Ltd., Hangzhou Zhuoyin Zhixin Technology Co., Ltd. and a partnership vehicle. Ruling: the operating entity, the wholly owning financing vehicle and the employee partnership are distinct; the brand is JoinAI / Zhuoyin Intelligence.
3. Investor attribution for the later round. Company and several reports call it an angel-plus round backed by Hangzhou Talent Fund Phase II and Hangzhou Runmiao Fund; one database labels it a Pre-A round dated 15 July 2026. Ruling: recorded as angel-plus with the Pre-A label noted as a media classification; no formal company announcement resolves the naming.
4. Timing of the later round. Sources place the announcement in July, on 11 September, or simply "recently" in 2026. Ruling: the round is accepted as closed in 2026; the precise date remains unconfirmed.
5. Customer list. Company materials name OpenAI, Google, SimplifyNext, Zhongke Chuangda and Zhonghao Xinying; investor accounts instead emphasize Huawei, Didi, Xiaomi, Ant Group, Ecovacs and Desay SV. Ruling: both lists are recorded as reported, with overlap only at SimplifyNext; neither is treated as a complete verified roster.
6. Cost-reduction claim. Earlier materials say under one tenth of cost and over ten times the speed; later materials say an 80 to 90 percent reduction in data-acquisition cost. Ruling: these are compatible phrasings of the same claim and are kept as a range, not a measured benchmark.
7. Model-improvement figure. Company POC tables show mIoU rising from 45.7 to 48.9 in one scenario; investor text summarizes mixing as 0.87 to 0.92 and a database claims over ten percent. Ruling: different scenarios and metrics, all first-party; recorded separately rather than collapsed.
8. Order and revenue claims. Investor materials cite over 20 million yuan in orders by May 2024 and exponential growth; a registry lists revenue of five million yuan in 2025 and eight employees in 2026. Ruling: orders and revenue are different measures from different sources and are not reconciled; both are flagged as unverified.
9. Terra versioning. Terra-1 is documented in detail on the site; Terra-3 appears in current marketing with object-level editing and spatial composition, with little public material on Terra-2 or the transition. Ruling: versions are recorded as named; capability continuity between them is not asserted.
10. Valuation. No authoritative public valuation exists, while some peers in the world-model space command billion-dollar valuations. Ruling: left blank; no inference made.
Risks and Due Diligence Gaps
The central risk is evidentiary. Nearly every performance and commercial claim — cost reduction, Learning Gain, customer roster, order value — originates with the company or its investors. The POC tables are unusually specific and internally consistent, which is a positive signal, but they remain self-selected scenarios without published protocols or independent reproduction. A prospective customer or investor should request third-party evaluation, named reference calls and signed contracts before treating the numbers as established.
A second risk is strategic positioning. JoinAI argues that data formats will change while the underlying experience a task requires will not, and uses this to justify an Experience Layer above specific formats. This is a sound framing but also a moving target: if world foundations from large platforms absorb scene generation, annotation and feedback into a single model stack, a thin integration layer could be squeezed. Conversely, if task-specific verification proves durable, the same position becomes defensible. The company's fate is tied to which of these industry paths dominates.
A third risk is organizational and financial opacity. Revenue, headcount and conversion from POC to recurring programs are not disclosed, and the financing narrative is assembled from media reports rather than formal announcements. The multi-entity structure, while normal for startups, complicates cap-table diligence. Finally, the founder's essays convey long-horizon ambition — building a model that can carry the real world, and a personal world for every user — that should be read as vision rather than near-term roadmap. Separating that vision from verifiable delivery is the essential analytical task for this company.
Overall Assessment
JoinAI is a technically credible early-stage company with a coherent and unusually well-articulated strategy: build the generative and management infrastructure that turns a model's capability gap into a verifiable Learning Gain. Its strengths are a fully in-house generative stack spanning Terra and Simulaix, the label-to-data trick, concrete POC numbers, a real customer list that includes global AI leaders, and backing from both a well-known early fund and two state-linked funds. Its weaknesses are the first-party nature of all evidence, undisclosed revenue and contracts, inconsistent financing labels and dates, and exposure to large platforms that could absorb its layer.
On this basis the five-dimension scores are: technology 7.5 (in-house models and a distinctive multi-output world; no independent benchmark), business 7.0 (real clients and paid POCs, but no revenue disclosure), funding 6.5 (angel plus backing, no valuation), product 7.5 (a coherent Terra-Simulaix-Chain stack plus standardized datasets) and market 7.0 (a recognized synthetic-data player in a still-forming category). The composite score is 7.1, reflecting a company whose engineering and narrative lead its verifiable commercial proof — a promising infrastructure sample worth tracking as the Physical AI data market takes shape.
References
- JoinAI official homepage — positioning, Physical AI training infrastructure, Terra-3 spatial composition and object-level editing (join-intelligence.com).
- JoinAI "About Us" — founder Kaiwen Yu, leadership team, 2022 Simulaix and 2023 Terra-1 milestones (join-intelligence.com/zh/about).
- JoinAI Terra-1 product page — world-model architecture, Transformer plus diffusion, multimodal inputs (join-intelligence.com/terra).
- JoinAI Simulaix product page — label-to-training-data, seven-to-three step pipeline, 80-90 percent cost reduction, three POC tables (join-intelligence.com/simulaix).
- WeChat — JoinAI completes angel-plus round, Experience Layer definition, customer list, Learning Gain (2026).
- PEdaily — company profile, 2021 registration, Angel Bay angel round, over 20 million yuan in orders (vc.pedaily.cn).
- Baidu Baike — registration history, cap table, Yu's indirect stake, revenue and headcount figures (baike.baidu.com).
- WeChat — dual state-linked funds, Terra/Simulaix/Chain stack, founder's Newcastle PhD (2026).
- Angel Bay Capital — founder interview on effect delivery, Experience Flywheel and competition (2026).
- Zhangtongshe — angel-plus round, Pre-A labeling and customer list (zhangtongshe.com).
- Dashu Cross-border — 20-plus deployment scenarios, 10-plus countries, national innovation contest award (10100.com).
Disclaimer: This report is compiled from public sources and company-published materials for research reference only. Performance figures, customer lists and financing details are first-party or media-reported and have not been independently verified; they do not constitute investment or procurement advice.
Related reports
- NVIDIA Deep Dive: Building the Full-Stack Physical-AI Infrastructure with PhysicsNeMo, Warp, Cosmos World Foundation Models and Omniverse
- World Labs: Fei-Fei Li's Spatial-Intelligence Unicorn, Acquired by AMD for $8.2B
- LiberAI (Jiangxian) Deep Dive: From RDT-1B to the Liber-0 World Action Model — a Tsinghua-spinoff physical-world-model company built on the UMI no-embodiment human-data + world-model paradigm
- Lightwheel Deep Dive: A Physical AI Data and Evaluation Infrastructure Startup Founded by a Former NVIDIA Simulation Lead, Seven Rounds Totalling Roughly CNY 2.65 Billion at a Valuation Above USD 2 Billion, Seats on Both the Newton and EgoVerse International Standards Committees, With Two Incompatible Accounts of Q1 Orders
- 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
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.