Jijia Vision (极佳视界 / GigaVision / GigaAI)

A world-model startup that raised 3.5 billion yuan in three months — but whose deployment evidence remains thin, whose benchmarks are self-reported, and whose claims of physical AGI leadership are running well ahead of independently verifiable results.

---

How to Read This Report

This report applies a four-tier evidence discipline to every material claim. Readers should weight assertions accordingly.

Sources are cited inline as bracketed numerals keyed to §14. Only URLs present in the research dossier are cited. Where the dossier is thin, that thinness is stated plainly.

---

01Executive Overview

Jijia Vision — registered as Beijing Jijia Vision Technology Co., Ltd., commercially branded GigaVision and GigaAI — is one of the most heavily capitalised embodied-intelligence startups to emerge from China in the current cycle. Founded in 2023 by Dr. Huang Guan, a Tsinghua-trained automation researcher with prior roles at Microsoft, Samsung, Horizon Robotics, and Jianzhirobot, the company has raised approximately 3.5 billion yuan across five disclosed rounds, with the final three rounds — totalling 3.5 billion yuan — completed within roughly three months in the first half of 2026 ¹³⁵⁹¹⁰. Investors include Huawei Hubble, Fortune Capital, Huakong Fund, and Lion City Capital ⁵¹¹⁰.

The company's stated thesis is that physical AI — robots and autonomous systems that act reliably in the unstructured physical world — requires a world-model foundation rather than a language-model foundation. Its core software stack comprises GigaWorld (a world-model platform with three published variants) and GigaBrain (an embodied vision-language-action model). Its flagship hardware product is the SeeLight S1, a wheeled-arm home robot approximately 1.6 metres tall with more than 20 degrees of freedom ⁷⁸. Supporting the stack are three data-collection hardware devices — Maker M01, U-01, and E-01 — and a stated ambition to accumulate one million hours of training data by end of 2026 ⁹.

The deployment record, as of the coverage date, consists of two confirmed pilots: 100 SeeLight S1 units placed in Wuhan Optics Valley talent apartments from 31 May 2026 for what the company explicitly describes as scenario testing, and a deployment at a FAW mould factory conducted in April 2026 in partnership with Alibaba Cloud ⁷⁹¹². A 1,000-robot three-year plan with Longsheng Technology in Wuxi has been announced but is a forward-looking commitment, not a current deployment ⁸. The company also claims more than 30 automaker and autonomous-vehicle company partners, though the nature and commercial terms of those relationships are not publicly disclosed ⁹.

The central tension in any analysis of Jijia Vision is the gap between the scale of capital raised and the thinness of independently verifiable deployment evidence. Benchmark claims — including first place on RoboChallenge, WorldArena, and RoboCasa365 — are vendor-reported and appear in trade press rather than in peer-reviewed venues or independently administered evaluations ¹¹⁷. The Wuhan trial is real but explicitly labelled as testing rather than productive deployment. The FAW factory deployment is described in press releases without third-party operational confirmation. This does not mean the technology is unserious — the affiliated research papers are substantive and the funding is genuine — but it does mean that the distance between the company's public narrative and its demonstrated, independently verified capability is considerable.

This report assesses the company's technology, commercial position, competitive context, and geopolitical exposure with the aim of giving a Max Robotics reader an accurate, evidence-graded picture of what Jijia Vision actually is at this stage of its development.

Latest news

• LG Electronics announces robotics division

  Biztoc.com·2026-06-30GENERAL
• Himax Strengthens 3D Sensing Portfolio with Launch of New iToF Depth Decoder ICs for Robotics and Intelligent Vision Applications

  GlobeNewswire·2026-06-30GENERAL
• Taiwan optics makers eye humanoid robot vision market as deployments accelerate

  Digitimes·2026-06-29GENERAL
• Esker Recognized as a Leader in Accounts Payable Automation

  GlobeNewswire·2026-06-23GENERAL
• Announcing VuePod: Turnkey Bin Picking Automation That Ships as a Product, Not a Project

  PRNewswire·2026-06-23GENERAL
• Mammotion LUBA 3 AWD 3000H AI Robot Mower for $2,239 + free shipping

  Dealnews.com·2026-06-23GENERAL
• Techman Robot Targets Southeast Asian Smart Manufacturing Markets at Thailand Automation Show

  Antaranews.com·2026-06-23GENERAL
• Doosan Robotics Unveils AI Palletizing Solution PalletizHD+ at Automate 2026

  PRNewswire·2026-06-22GENERAL

---

02The Jijia Vision Story

Founding and Founder

Jijia Vision was founded in 2023 ¹⁵. The founding year places it squarely in the second wave of Chinese embodied-intelligence startups — after the early movers such as Unitree and Agility-adjacent ventures, but contemporaneous with a cohort of model-first companies that emerged as large language models demonstrated that foundation-model approaches could transfer to physical systems.

The founder and CEO, Dr. Huang Guan, holds a PhD from Tsinghua University's Department of Automation ⁵. His career trajectory is instructive: Microsoft and Samsung provided exposure to large-scale software and consumer hardware; Horizon Robotics, one of China's leading automotive AI chip companies, gave him grounding in vision perception at the edge; and Jianzhirobot, a Chinese humanoid startup, provided direct embodied-robotics experience as a partner and algorithm vice-president ⁵. This combination — perception research, edge inference, and robotics — maps directly onto the technical problem Jijia Vision is attempting to solve. The company's emphasis on 3D geometric understanding over language-model backbones (discussed in §4) is consistent with a founder who spent years on vision perception rather than NLP.

The company's Chinese name, 极佳视界, translates approximately as "excellent vision world" or "supreme visual horizon." The English brand GigaVision emphasises the visual-perception lineage; GigaAI is used in consumer-facing contexts, particularly around the SeeLight S1 ⁷.

The Funding Trajectory

The funding history is the most striking single fact about Jijia Vision. The sequence, as reconstructed from multiple corroborating sources, is as follows ^1591015:

The acceleration from A2 (200 million yuan) to Pre-B (1 billion yuan) to B1 (1.5 billion yuan) to B2 (1 billion yuan) within approximately three months is unusual even by the standards of the current Chinese AI investment cycle ³⁶¹⁰. The B2 round was reported as oversubscribed ¹⁰. Valuation at any round is not publicly disclosed ¹⁰.

Huawei Hubble's participation at Series A1 is strategically significant and warrants separate treatment in §10. Huawei Hubble is the investment arm of Huawei Technologies and has a pattern of taking early stakes in Chinese semiconductor, AI, and robotics companies that may eventually integrate with Huawei's hardware and software ecosystem. Its presence signals that Jijia Vision's technology is considered strategically relevant at the platform level, not merely as a standalone product company.

Lion City Capital, the B2 lead, is a Singapore-based fund, which introduces a cross-border capital dimension that has implications for the company's potential international positioning and for its exposure to technology-export scrutiny (see §10) ¹⁰.

The Stated Mission

The company describes its mission as building "physical AGI" — artificial general intelligence that acts directly on the physical world rather than generating text or images ¹². Dr. Huang has stated publicly that physical AGI will "directly act on the real physical world" and that the world-model approach is the correct technical route to achieve this ¹²⁵. The language of AGI in a robotics context is a marketing and fundraising frame as much as a technical claim; the editorial inference here is that the company is positioning itself as a foundational-infrastructure play rather than a single-product hardware company, which justifies the capital intensity of its roadmap and the breadth of its stated ambitions.

The company's self-description as pursuing "the ultimate technical route" for physical AI ⁵ is a COMPANY CLAIM. No independent technical authority has endorsed this framing. The research papers affiliated with the company (discussed in §5) do make substantive technical arguments for the world-model approach, but the leap from "technically well-argued" to "ultimate route" is not supported by the current evidence base.

Early Operational History

Beyond the funding narrative, the operational history of Jijia Vision before 2026 is not well documented in the public record. The company appears to have spent 2023 and 2024 in research and development, with the DriveDreamer series — a set of world models for autonomous driving — as an early output ⁹. The transition from autonomous-driving world models to embodied robotics world models is described as a deliberate strategic expansion rather than a pivot, with the company arguing that the underlying world-model architecture is domain-agnostic ⁹. This claim is technically plausible but remains a COMPANY CLAIM; the degree to which autonomous-driving world-model training transfers to manipulation tasks is an open research question.

The first major public deployment milestone — the Wuhan apartment trial — was announced on 31 May 2026 ⁷. The FAW factory deployment was announced in April 2026 ⁹. Both are recent enough that no independent operational assessment is available in the supplied evidence base.

---

03Product Portfolio: What Jijia Vision Actually Sells

Jijia Vision's product portfolio spans three layers: a world-model software platform, an embodied VLA model, and physical hardware including a consumer-facing robot and data-collection devices. The company presents these as an integrated stack rather than separable products, which is consistent with its platform-company positioning.

GigaWorld: The World-Model Platform

GigaWorld is described as a world-model platform with three published variants ⁵⁹:

• GigaWorld-0: Described by the company as "the world's first systematically introduced embodied world model." This is a COMPANY CLAIM. No independent source in the dossier confirms the "world's first" designation, and the autonomous-driving world-model field (including Wayve's GAIA-1, NVIDIA's DriveDreamer antecedents, and others) predates this claim in adjacent domains.
• GigaWorld-1: The second-generation world model. The company claims it achieved first place on the WorldArena benchmark with a score of 62.34, described as the first model to exceed 60 on that leaderboard ¹¹. This is a COMPANY CLAIM; the independence of the WorldArena benchmark from the company is not established in the supplied facts.
• GigaWorld-Policy: A policy-generation variant of the world model. The company claims first place on the RoboCasa365 benchmark, ahead of NVIDIA GR00T N1.5 and Physical Intelligence's π0.5 ¹¹. Again, this is a COMPANY CLAIM with no independent verification in the dossier.

The platform's architecture is described as a "dual-model" design comprising a World Generation component and a World Action component ⁵. The World Generation component produces predictions about future world states; the World Action component translates those predictions into robot control policies. This architecture is consistent with the broader academic literature on model-based reinforcement learning and world models, and the affiliated research papers (§5) provide technical elaboration.

GigaBrain: The Embodied VLA Foundation Model

GigaBrain-0 is the company's embodied vision-language-action foundation model ¹⁵. The company claims it achieved first place on the RoboChallenge benchmark with a 51.67% task success rate, approximately 10 percentage points ahead of π0.5 ¹¹. This is a COMPANY CLAIM.

The company describes GigaBrain as operating at "the leading level in China" for embodied intelligence ¹. The claim of 85% average success rate across grasping, assembly, and sorting tasks is also vendor-reported ⁷. No independent third-party evaluation of GigaBrain-0 is present in the supplied evidence base.

SeeLight S1: The Home Robot

The SeeLight S1 (拾光S1, literally "gathering light S1") is the company's flagship hardware product and the vehicle for its consumer market entry ⁷⁸. Confirmed specifications from vendor sources:

The choice of a wheeled chassis rather than a legged platform is a deliberate engineering decision that prioritises stability and indoor navigability over stair-climbing capability. This is consistent with the home-service use case but limits deployment environments. The multi-fingered gripper is the critical manipulation interface; the quality of that gripper's dexterity in unstructured household environments is the central unknown in assessing the S1's practical utility.

The Wuhan Optics Valley deployment of 100 S1 units from 31 May 2026 is the primary evidence of the product's real-world existence ⁷. The deployment is explicitly described as "scenario testing" — meaning it is a structured pilot to gather data and validate performance, not a commercial rollout. This distinction matters: the company's press materials tend to describe the deployment in terms that imply operational readiness, while the underlying characterisation is of a data-collection and validation exercise.

Pricing for the SeeLight S1 is not publicly disclosed.

Data-Collection Hardware: Maker M01, U-01, E-01

Three data-collection hardware devices are listed in vendor materials ⁹:

• Maker M01: Robot-mounted data collection hardware
• U-01: Handheld data collection device
• E-01: Ego (first-person) data collection device

These devices serve the company's data strategy (discussed in §4) rather than being standalone commercial products. Their specifications, pricing, and availability are not publicly disclosed. Their existence reflects the company's recognition that proprietary training data is a competitive moat — a sound strategic insight in the current embodied-AI landscape.

DriveDreamer: The Autonomous Driving Heritage

The company references a "DriveDreamer series" of world models for autonomous driving, described as having "achieved large-scale mass production and implementation" ⁹. This is a COMPANY CLAIM. The actual scale of DriveDreamer deployment is not independently verifiable from the supplied evidence. The claim of "large-scale mass production" is inconsistent with the company's 2023 founding date and the limited deployment evidence available; the editorial inference is that this language describes integration with automotive partners' development pipelines rather than mass-market consumer deployment.

The 30-plus automaker and autonomous-vehicle company partners claimed by the company ⁹ are not named individually in the supplied evidence, making independent verification impossible. The FAW factory deployment (April 2026, with Alibaba Cloud) is the only named industrial partner confirmed across multiple sources ⁹¹².

Products & versions

SeeLight S1 (拾光S1)

Wheeled-arm home robot ~1.6 m tall with 20+ DOF, omni-directional chassis, kg-level arm payload, and multi-fingered grippers; piloted in Wuhan Optics Valley residential apartments from May 2026.

GigaWorld Platform

World-model platform comprising GigaWorld-0, GigaWorld-1, and GigaWorld-Policy; generates synthetic training data and action policies for embodied robots; vendor-claimed #1 on WorldArena and RoboCasa365 benchmarks.

GigaBrain-0

Embodied vision-language-action (VLA) foundation model; vendor-claimed #1 on RoboChallenge with 51.67% task success rate, approximately 10 percentage points ahead of π0.5.

Maker M01

Robot-mounted hardware device for collecting embodied manipulation training data as part of the company's five-layer Data Pyramid strategy.

U-01

Handheld data collection device for capturing human manipulation demonstrations to feed the GigaWorld/GigaBrain training pipeline.

E-01

Ego-perspective data collection device used to gather first-person manipulation data for embodied AI training.

---

04Technology Stack: Strengths and the Work That Remains

The Core Architectural Argument

Jijia Vision's central technical claim is that existing approaches to embodied AI — specifically those that adapt large language models or video-generation models as robot-control backbones — are architecturally misaligned with the requirements of physical manipulation ¹⁸. The argument, elaborated in the VGA paper (§5), is that robotic manipulation is fundamentally a vision-to-geometry mapping problem: the robot must understand the 3D geometric structure of its environment and the objects within it, not merely generate plausible-looking video frames or produce language tokens ¹⁸. Language and video models, trained on 2D internet data, lack the geometric priors necessary for reliable physical interaction.

This is a substantive technical argument with genuine support in the robotics research community. The difficulty of transferring 2D vision-language models to 3D manipulation tasks is well-documented. The question is whether Jijia Vision's specific architectural response — the Progressive Volumetric Modulation approach in VGA, the Joint Visuomotor Gating in FutureVLA, and the triple-system architecture in TriVLA — constitutes a decisive advance or one contribution among many in a rapidly moving field.

The Three Research Models

Three research models are attributed to researchers affiliated with the company or its academic partners:

VGA (Vision-Geometry-Action): Uses a pretrained 3D world-model backbone with Progressive Volumetric Modulation to inject geometric structure into the action-prediction pipeline ¹⁸. The core claim is that treating manipulation as a geometry-mapping problem rather than a language or video problem produces better generalisation to novel objects and environments. The paper is available on arXiv ¹⁸.

FutureVLA: Introduces a Joint Visuomotor Predictive Architecture with decoupled visual and motor streams and a Joint Visuomotor Gating mechanism ¹⁹. The key insight is that predicting future visual states and predicting future motor commands are related but not identical problems, and that decoupling them while allowing gated interaction improves both. The paper is on arXiv ¹⁹.

TriVLA: A triple-system architecture combining a vision-language model, a dynamics perception module, and a motor policy module, operating at approximately 36 Hz ²⁰. The 36 Hz inference rate is significant: it is fast enough for real-time reactive control in most manipulation scenarios, though not for the fastest dynamic tasks. The paper is on arXiv ²⁰.

A fourth paper, on a dual-layer VLM-guided precision manipulation framework, is also in the affiliated research corpus ²¹.

Strengths

Geometric grounding: The emphasis on 3D geometric understanding is technically well-motivated. The VGA paper's argument that video and language backbones are misaligned with manipulation requirements is consistent with findings from multiple independent research groups.

Real-time inference: The 36 Hz operating rate claimed for TriVLA ²⁰ is a practical engineering achievement if verified. Many academic VLA systems operate at much lower frequencies, making them unsuitable for reactive manipulation.

Data strategy coherence: The Five-Layer Data Pyramid and the suite of data-collection hardware (Maker M01, U-01, E-01) reflect a coherent understanding that data quality and diversity, not just model architecture, determine real-world performance ⁹. The target of one million hours of training data by end of 2026 is ambitious; whether it is achievable is unknown.

World-model integration: The dual World Generation / World Action architecture provides a principled framework for using world-model predictions to guide policy, rather than treating world modelling and policy learning as separate problems. This is architecturally sound.

The Work That Remains

Benchmark independence: All performance claims — RoboChallenge, WorldArena, RoboCasa365 — are vendor-reported or appear in company-adjacent press releases ¹¹⁷. The independence of these benchmarks from the company is not established. Until results are reproduced on independently administered benchmarks (e.g., LIBERO, RLBench, or benchmarks administered by academic groups with no financial relationship to the company), the claimed performance advantages over π0.5 and GR00T N1.5 cannot be accepted as verified.

Sim-to-real transfer: The degree to which world-model-generated synthetic data transfers to real-world robot performance is the central unresolved question in the field. The company's claim that world-model-generated data can substitute for real-machine data at scale is plausible in principle but unproven at the scale claimed.

Dexterous manipulation: The SeeLight S1's multi-fingered gripper is described in general terms; the specific dexterity capabilities — precision grasping of small objects, handling of deformable materials, manipulation in cluttered environments — are not documented in the supplied evidence. These are precisely the tasks where current VLA systems most frequently fail.

Generalisation beyond trained scenarios: The Wuhan apartment trial is explicitly a scenario-testing exercise, implying that the robot is being evaluated on a defined set of tasks rather than deployed in fully open-ended household environments. The gap between scenario-tested performance and genuine open-world generalisation is the hardest problem in embodied AI.

Inference at the edge: The TriVLA 36 Hz claim does not specify whether this rate is achieved on-device or via cloud inference. For a home robot operating in environments with variable connectivity, on-device inference capability is a practical requirement. This is not publicly disclosed.

Gripper and manipulation hardware quality: No independent teardown, component-level analysis, or third-party mechanical assessment of the SeeLight S1 is present in the supplied evidence. The quality of the physical manipulation hardware is as important as the software stack for real-world task success.

The Data Bottleneck

The company's own materials acknowledge that data scarcity, high cost of real-machine data collection, and the misalignment of existing video backbones with 3D geometric requirements are the primary challenges facing the field ⁹¹⁸. This is an honest acknowledgement of genuine problems. The proposed solution — a Five-Layer Data Pyramid combining real-machine data, internet human manipulation data, and world-model-generated synthetic data — is directionally correct but faces the same fundamental challenge as every other embodied-AI data strategy: the distribution of synthetic data may not match the distribution of real-world manipulation failures, which are precisely the cases the model most needs to learn from.

The 10x inference speed and training efficiency claim relative to prior methods ⁷ is a COMPANY CLAIM with no independent benchmark or ablation study cited in the supplied evidence.

---

05Research, Papers, Authors and Labs

The Research Corpus

Four arXiv preprints are directly relevant to Jijia Vision's technical claims, based on the supplied evidence. All are preprints; none has been confirmed as peer-reviewed and published in a major venue as of the coverage date.

Institutional Affiliations

The research papers are affiliated with Sun Yat-sen University, X-Era AI Lab, AMAP (the mapping division of Alibaba), Guangdong University of Technology, and the Guangdong Key Laboratory of Big Data Analysis and Processing ¹⁸. The presence of AMAP/Alibaba affiliation is notable: it suggests a research relationship with Alibaba's technical organisation that may be connected to the Alibaba Cloud partnership on the FAW factory deployment ⁹.

The X-Era AI Lab affiliation is the most directly company-adjacent; this appears to be a research lab associated with Jijia Vision's technical team, though the precise organisational relationship is not publicly documented in the supplied evidence.

What the Papers Establish and What They Do Not

The papers establish that the affiliated research team is engaged with substantive technical problems in embodied AI, that their architectural proposals are well-motivated and internally coherent, and that they are producing work at a pace consistent with a well-resourced research organisation. The VGA paper in particular makes a clear and falsifiable argument — that 3D geometric backbones outperform language and video backbones for manipulation — that can be tested by independent researchers.

What the papers do not establish: that the described architectures are deployed in the SeeLight S1 or GigaBrain-0 as described; that the performance claims in the papers translate to the benchmark claims in the press releases; or that the approaches generalise beyond the specific task sets evaluated in the papers. The relationship between the research papers and the commercial product claims is asserted by the company but not independently documented.

Gaps in the Research Record

No peer-reviewed journal or top-tier conference publications (NeurIPS, ICRA, ICLR, CoRL, RSS) from Jijia Vision or its directly affiliated researchers are confirmed in the supplied evidence. All four papers are arXiv preprints. This is not unusual for a company at this stage — the publication cycle for top venues is long — but it means the work has not yet passed external peer review. The absence of code releases or public model weights also limits independent reproducibility assessment.

Company-linked papers

This module is being compiled — no data to show yet.

Authors & labs

This module is being compiled — no data to show yet.

Code & simulation

This module is being compiled — no data to show yet.

Datasets & benchmarks

This module is being compiled — no data to show yet.

---

06Media Evidence Library: What the Videos Prove

The Available Video Evidence

The supplied research dossier contains six video sources. Of these, three are confirmed misattributions — Apple Vision Pro teardown and review videos ²²²³²⁴ — and two others are clearly unrelated to Jijia Vision (a JIYI drone radar unboxing ²⁵ and a Gemma 4 photo-coaching app demo ²⁷). One further video ²⁶ concerns a portable gaming monitor review. None of these six videos contains footage of Jijia Vision products or technology.

This is a significant evidential gap. For a company that has raised 3.5 billion yuan, deployed 100 robots in Wuhan, and claims benchmark leadership in embodied AI, the absence of independently sourced, third-party video documentation of the SeeLight S1 operating in real environments is notable. The company almost certainly has produced promotional video content — Chinese robotics startups at this funding level invariably do — but no such content appears in the supplied dossier in a form that can be independently assessed.

What Can and Cannot Be Concluded from Promotional Video

Even if promotional video of the SeeLight S1 were available, the appropriate analytical standard is strict. A choreographed demonstration video — the standard format for Chinese robotics startup product reveals — proves the following: that the hardware exists in at least one functional unit; that the robot can perform the specific demonstrated tasks under the specific demonstrated conditions; and that the company has sufficient engineering capability to produce a working prototype. It does not prove: autonomous operation in unstructured environments; generalisation to tasks not shown; reliability at scale; or readiness for unsupervised deployment.

The Wuhan apartment trial, if documented on video, would be more informative than a studio demonstration — but only if the video showed unscripted interactions, failure modes, and recovery behaviours, rather than a curated selection of successful task completions.

The Absence of Independent User Documentation

No independent user reports, social media documentation from Wuhan apartment residents, or third-party journalist visits to the Wuhan deployment are present in the supplied evidence. The 100-unit trial began on 31 May 2026 ⁷, giving approximately three to four weeks before the coverage date for such documentation to emerge. Its absence does not prove the trial is not occurring, but it does mean that the only account of the trial's progress is the company's own.

The editorial inference is that the Wuhan trial is real — the announcement is corroborated across multiple news sources ⁷⁹¹² — but that its operational character (what tasks the robots are actually performing, how often they succeed, what failure modes have been observed, what level of human supervision is required) is entirely undocumented from independent sources.

Media library

Tesla Optimus Gen 3: Elon Musk JUST Revealed a NEW Humanoid Robot To Win The AI Race!

YouTube

#AUTOMATE2025 @VisionNavRobotics Forklift #AMR loading/unloading #robotics at @automateshow 2025

YouTube

Figure's Robots Conquer Logistics Sorting Packages with AI Vision!

YouTube

Gausium | Smarter and Safer Autonomous Cleaning Powered by AI Vision

YouTube

Best Window Cleaning Robots 2025 (watch this before you buy)

YouTube

Robots Replacing Humans? The Future of Work EXPLAINED!

YouTube

Human robot interaction

YouTube

Envision Robotics - LEGO Spike Prime Robot Alligator

YouTube

Boston Dynamics NEW Atlas AI Robot Just Broke Another Human Skill Barrier!

YouTube

Machine Vision (Based on Halcon)

Bilibili385k views

Robotics: Modeling, Control, and Vision - Huazhong University of Science and Technology

Bilibili343k views

Quick Learn VisionPro Series Tutorial (Industrial Machine Vision Applications)

Bilibili313k views

Hikvision AGV/AMR Intelligent Mobile Robot Basics Knowledge Sharing Full Version

Bilibili118k views

Hikrobot Machine Vision Certified Engineer (Beginner) Tutorial! Full Version!

Bilibili91k views

Machine Vision - AI Harvesting Robot

Bilibili75k views

Hikvision AGV Cart Debugging, Forklift Drivers Facing Elimination!

Bilibili49k views

Six Referenced Fixed-Wing Drones Have Clear Division of Labor, High-End Autonomous Air Combat Drones Deployed, U.S. Counterpart Project Increment 2 Still in Planning Stage.

Bilibili36k views

Bill Gates: AI is the biggest technical achievement I've ever seen in my lifetime - CNBC Television

Bilibili25k views

---

07Commercial Reality

What Is Actually Deployed

As of 25 June 2026, the confirmed deployment record for Jijia Vision consists of two pilots:

The Wuhan deployment is the more significant of the two from a commercial standpoint, as it represents the company's entry into the home-service market. However, "scenario testing" is a precise and important qualifier: it means the robots are being evaluated against defined task scenarios, not deployed to provide unsupervised household services. The residents of the Wuhan apartments are, in effect, participants in a structured pilot study, not customers receiving a commercial service.

The FAW factory deployment is described as "full-process" ⁹, which implies end-to-end integration into a manufacturing workflow. However, the scale of this deployment — number of robots, specific tasks, throughput metrics, uptime — is not disclosed. The involvement of Alibaba Cloud suggests a cloud-inference architecture rather than fully on-device operation, which has implications for latency, connectivity dependency, and data sovereignty.

Forward-Looking Commercial Commitments

Beyond the two confirmed pilots, the company has announced:

• A 1,000-robot, three-year plan with Longsheng Technology in Wuxi ⁸. This is a forward-looking commitment, not a current deployment. The terms, pricing, and milestones of this agreement are not publicly disclosed.
• A broader Wuhan seed-user programme planned for 2027 ⁷. This is a stated intention, not a confirmed commercial arrangement.
• More than 30 automaker and autonomous-vehicle company partners ⁹. These are unnamed and the nature of the relationships (paid contracts, development partnerships, letters of intent, or informal collaborations) is not disclosed.

Revenue and Unit Economics

Revenue figures are not publicly disclosed. Pricing for the SeeLight S1 is not publicly disclosed. The cost structure of the GigaWorld and GigaBrain platforms — whether sold as SaaS, licensed per-robot, or bundled with hardware — is not publicly disclosed. Given the company's stage (Series B2, pilot deployments), the editorial inference is that revenue is minimal relative to the capital raised, and that the company is operating on a venture-funded growth model with commercial revenue as a future rather than current primary income source.

The 3.5 billion yuan raised in approximately three months ³¹⁰ implies a burn rate and capital deployment plan of considerable scale. The stated uses of capital include data acquisition (targeting one million hours of training data by end of 2026), hardware manufacturing scale-up, and continued model development ⁹¹². At a rough estimate, one million hours of robot training data at even modest collection costs represents a very large expenditure; the world-model-generated synthetic data strategy is partly motivated by the need to reduce this cost ⁹.

The Claim-vs-Reality Gap in Commercial Language

The company's press materials use language that consistently overstates the current commercial position. The description of DriveDreamer as having "achieved large-scale mass production and implementation" ⁹ is inconsistent with the evidence of a company founded in 2023 with its first confirmed hardware deployment in April 2026. The description of the Wuhan trial in some press coverage as a "deployment" rather than a "scenario test" elides an important distinction. The 30-plus partner claim, without named partners or disclosed contract terms, is a standard Chinese startup press-release convention that provides little commercial signal.

This pattern of language — common across the Chinese AI startup ecosystem — does not necessarily indicate bad faith. It reflects a fundraising and partnership-development culture in which forward-looking ambition is presented alongside current reality without always clearly distinguishing the two. For an analyst or investor, the discipline of separating what is deployed from what is planned, and what is contracted from what is announced, is essential.

The Competitive Funding Context

The speed and scale of Jijia Vision's fundraising — 3.5 billion yuan in three months — places it among the most aggressively capitalised embodied-AI startups in China in 2026. For context, this is approximately 480 million US dollars at current exchange rates, raised by a company with a deployment record of 100 robots in a pilot and one factory integration. The capital is real; the deployment scale is not yet commensurate with it. The editorial inference is that investors are betting on the world-model thesis and the team's technical credentials rather than on demonstrated commercial traction — a rational bet in a market where the technology is developing rapidly and first-mover advantages in data and model quality may be durable, but a bet that carries substantial execution risk.

Customers & deployments

Wuhan Optics Valley Talent ApartmentsResidential Property

100 SeeLight S1 units entered Wuhan Optics Valley talent apartments from May 31, 2026 for scenario testing, with a broader Wuhan seed-user program planned for 2027.

FAW Mould FactoryAutomotive Manufacturer

Full-process deployment of Jijia Vision robots at a FAW mould factory in April 2026, conducted in partnership with Alibaba Cloud.

Longsheng TechnologyIndustrial / Manufacturing

Signed a 1,000-robot, 3-year deployment plan with Jijia Vision in Wuxi; forward-looking target, not yet fully executed.

08Markets and Use Cases

Jijia Vision has staked out two primary market verticals — domestic service robotics and industrial manipulation — with autonomous driving as a stated third vector through its world-model lineage. The company's public positioning treats these not as separate product lines but as a single platform play: the same GigaWorld and GigaBrain stack, retrained or fine-tuned for different task domains. Whether that architectural ambition translates into genuine cross-domain transferability remains unproven at the current pilot stage.

Domestic Service Robotics

The SeeLight S1 is the company's most visible commercial bet. The target customer is not the mass consumer market — not yet — but rather institutional residential operators: talent apartment complexes, managed housing estates, and eventually elder-care facilities. The Wuhan Optics Valley deployment, which placed 100 units into talent apartments from 31 May 2026, is explicitly framed as "scenario testing" rather than commercial sale ⁷. This is an important distinction. The units are in the field to generate training data and surface failure modes, not to deliver a paid service at scale.

The stated 2027 timeline for a broader Wuhan seed-user programme suggests the company itself does not expect consumer-ready performance before that date ⁷. The elder-care angle — mentioned in the B2 round coverage — is strategically sensible given China's demographic trajectory, but it is the highest-stakes environment for reliability: a robot that drops a medication bottle or fails to detect a fall is a liability event, not merely an inconvenience ⁸.

The home-robot market in China is nascent and contested. Penetration of any wheeled-arm home robot beyond technology enthusiasts and institutional pilots remains negligible industry-wide. Jijia Vision is not alone in targeting this space; it is competing against better-capitalised incumbents and better-known brands. The company's differentiation argument rests on the world-model approach producing more generalisable manipulation behaviour — a claim that is architecturally coherent but empirically unverified in independent testing.

Industrial Manipulation

The FAW Mould factory deployment, executed jointly with Alibaba Cloud in April 2026, represents the company's most substantive industrial footprint to date ⁹. Mould-handling in automotive manufacturing involves heavy, geometrically complex parts with tight tolerances — a genuinely demanding manipulation context. However, the public record does not specify which tasks the robot performs autonomously versus which are supervised or teleoperated, nor does it provide throughput, uptime, or defect-rate data.

The 1,000-robot, three-year plan with Longsheng Technology in Wuxi is a forward-looking commercial agreement, not a current deployment ³. It signals intent and provides a revenue pipeline if fulfilled, but three-year robotics supply agreements in China's manufacturing sector have a mixed completion record, particularly when the supplier is a pre-revenue startup.

The claim of 30-plus automaker and autonomous-vehicle company partners is vendor-reported and unverified ⁵. In the Chinese automotive-technology ecosystem, "partner" can mean anything from a signed letter of intent to a paid production contract. Without named customers confirming active deployments, this figure should be treated as a pipeline indicator rather than a commercial fact.

Autonomous Driving Data Infrastructure

Jijia Vision's DriveDreamer lineage — the world-model work that predates the embodied robotics pivot — positions the company as a potential supplier of synthetic training data and simulation environments to autonomous-driving developers. The company claims "large-scale mass production and implementation" for the DriveDreamer series ⁵, but independent evidence supports only early-stage pilot relationships. This vertical is less prominent in recent communications, suggesting the company's commercial centre of gravity has shifted toward embodied robotics, where the funding narrative is currently stronger.

Use-Case Plausibility Assessment

The table above reflects editorial inference from available evidence, not vendor claims. The honest read is that Jijia Vision is a platform-stage company with credible technology hypotheses and two real but modest deployments, operating in markets that will take three to five years to reach meaningful volume.

---

09Competitive Landscape

Jijia Vision enters a competitive field that is simultaneously crowded at the venture level and thin at the proven-deployment level. The world-model-for-robotics thesis is not unique to this company; several well-funded Chinese and international players are pursuing structurally similar approaches. What differentiates Jijia Vision — if the claims hold — is the depth of the world-model integration into both the perception backbone and the policy layer, rather than treating world models as a bolt-on data-augmentation tool.

Chinese Domestic Competitors

Unitree Robotics is the most commercially mature Chinese humanoid and quadruped manufacturer, with documented hardware sales and a developer ecosystem. Unitree's competitive advantage is hardware iteration speed and price competitiveness; its software stack is less differentiated on the world-model axis. Jijia Vision's wheeled-arm form factor does not directly compete with Unitree's bipedal G1 or quadruped Go2, but both companies are targeting the industrial manipulation market.

Agibot (稀宝) and Galbot are Chinese humanoid startups with comparable funding vintages and similar VLA-model ambitions. Agibot has disclosed more hardware deployment data and has a clearer humanoid form-factor strategy. Galbot's wheeled-arm approach is the most direct architectural parallel to the SeeLight S1.

DJI and UBTECH represent established Chinese robotics capital with longer track records. Neither is pursuing the world-model-first software strategy in the same explicit framing as Jijia Vision, but both have the manufacturing and distribution infrastructure that Jijia Vision lacks.

Zhiyuan Robotics (AgiX) and Fourier Intelligence are pursuing humanoid form factors with industrial deployment focus. Their competitive threat to Jijia Vision is indirect but real: if humanoid robots achieve sufficient dexterity and cost reduction, the wheeled-arm form factor loses its ergonomic-compromise argument.

International Competitors

Physical Intelligence (pi) is the most directly comparable international competitor on the software-model axis. The GigaBrain-0 benchmark claim of outperforming pi's pi0.5 by approximately 10 percentage points on RoboChallenge is the company's most pointed competitive assertion ¹. This claim is vendor-reported and unverified. Pi has published peer-reviewed work and has disclosed named enterprise pilots; Jijia Vision has neither at the time of writing.

Figure AI and 1X Technologies are pursuing humanoid platforms with proprietary VLA models. Their funding scale is comparable to Jijia Vision's post-B2 position, but their hardware is more mature.

Boston Dynamics (Hyundai) remains the benchmark for hardware reliability and real-world deployment longevity, though its software AI stack is less differentiated on the generative-model axis.

Google DeepMind (RT-2, RT-X) and Nvidia (GR00T N1.5) represent the hyperscaler threat: foundation models trained on vastly larger compute budgets. The GigaWorld-Policy claim of outperforming Nvidia GR00T N1.5 on RoboCasa365 is the most audacious competitive assertion in the dossier ³. If true, it would represent a significant efficiency advantage. If false or cherry-picked, it is a reputational liability.

Competitive Positioning Summary

The competitive picture is one of a well-funded but early-stage entrant making aggressive benchmark claims against more established players. The claims may prove accurate as independent evaluations emerge, but at present Jijia Vision's competitive position rests more on capital raised and narrative than on verified performance differentiation.

Competitive comparison

Robot	Maker	Autonomy	Conf.
iRobot Roomba Combo 10 Max	iRobot	Autonomous	0.90
Mobile ALOHA (Stanford)	Stanford University	Teleoperated	0.90
1X NEO	1X Technologies	Remote-Assisted	0.90

---

10Geopolitical Context and Constraints

Jijia Vision operates at the intersection of three geopolitically sensitive domains: artificial intelligence, robotics, and autonomous systems. Each carries its own regulatory and export-control exposure; the combination amplifies the risk profile for any non-Chinese investor or customer considering engagement.

The China AI Investment Surge

The 3.5 billion yuan raised in approximately three months in 2026 is not an isolated data point. It reflects a broader pattern of accelerated Chinese venture capital deployment into physical AI, driven partly by the domestic policy priority of achieving leadership in robotics and AI by 2030, and partly by the competitive anxiety generated by international developments in large language and action models. Huawei Hubble's participation in the Series A1 is particularly significant ⁵. Huawei Hubble is the strategic investment arm of Huawei Technologies, a company that is itself subject to US export controls on semiconductors and software. Hubble's portfolio companies benefit from Huawei's ecosystem — including Ascend AI chips and the HarmonyOS platform — but they also inherit Huawei's geopolitical exposure.

Semiconductor and Compute Dependencies

Jijia Vision's world-model training at scale requires substantial GPU compute. The company has not publicly disclosed its compute infrastructure, but Chinese AI startups in 2025-2026 face a constrained environment: Nvidia H100 and H800 chips are subject to US export controls, and domestic alternatives (Huawei Ascend 910B/C, Cambricon, Biren) lag in software ecosystem maturity and in some cases raw throughput. If Jijia Vision is training on Ascend hardware, its benchmark comparisons against models trained on Nvidia infrastructure carry an implicit asterisk: the training regime, data pipeline, and optimisation stack may differ in ways that affect reproducibility.

This is not a company-specific criticism — it applies to the entire Chinese AI training ecosystem — but it is a material unknown for any technical due-diligence exercise.

Export Controls and Dual-Use Classification

Robotic manipulation systems with advanced perception and autonomous decision-making capabilities sit in a grey zone under both Chinese and international export-control frameworks. China's own export-control law (2020) and the subsequent AI governance regulations create obligations on Chinese AI companies around data security and cross-border data transfer. For Jijia Vision, which is collecting real-world manipulation data from residential apartments and automotive factories, the data governance question is non-trivial: who owns the data generated by the SeeLight S1 in a Wuhan apartment, and under what conditions can it be used for model training?

The company's stated target of 1 million hours of training data by end of 2026 ⁹ implies large-scale data collection from real-world environments. The regulatory framework governing this collection — particularly in residential settings — is not addressed in any public document in the dossier.

The Wuhan Nexus

Wuhan is not an arbitrary deployment location. It is the capital of Hubei Province and a designated hub for China's intelligent manufacturing and new-energy-vehicle industries. The Wuhan municipal government has been an active supporter of robotics and AI companies, and the Optics Valley (光谷) district where the SeeLight S1 is being trialled is a government-backed technology zone. This suggests the Wuhan deployment may benefit from municipal subsidy or facilitation, which would affect the economics of the pilot in ways that do not generalise to commercial deployments elsewhere.

Talent and Brain-Drain Dynamics

Dr. Huang Guan's trajectory — Tsinghua PhD, Microsoft, Samsung, Horizon Robotics, Jianzhirobot — is representative of a generation of Chinese AI researchers who built careers partly in international technology companies before returning to found domestic startups ¹. This pattern is well-documented and reflects both the maturation of China's domestic AI ecosystem and the increasing difficulty of senior Chinese researchers obtaining or retaining US work authorisation. The research affiliations of Jijia Vision's published papers — Sun Yat-sen University, Guangdong University of Technology — suggest the company is drawing on southern China's academic talent pool, which is less internationally prominent than Beijing or Shanghai but has produced credible robotics research.

Geopolitical Risk for Non-Chinese Partners

Any non-Chinese company considering a technology partnership, data-sharing arrangement, or supply relationship with Jijia Vision should conduct careful legal review under the following frameworks: US Entity List screening (Huawei Hubble's involvement warrants attention), EU AI Act compliance (if deploying in European residential settings), and domestic data-protection law in any jurisdiction where the robot collects environmental data. These are not hypothetical concerns; they are standard due-diligence requirements for any Chinese physical-AI company with this investor profile.

---

11The Hype, the Real and the Ugly

Jijia Vision's public communications exhibit a pattern common to well-funded Chinese AI startups: a mixture of genuine technical substance, aggressive benchmark claims, and marketing language that conflates aspiration with achievement. Separating these layers is the core analytical task.

What Appears Genuine

The founding team's credentials are verifiable and strong. Dr. Huang Guan's background at Horizon Robotics — where he led vision perception — is directly relevant to the company's technical thesis ¹. The research papers (VGA, FutureVLA, TriVLA) are published on arXiv with institutional affiliations and represent coherent, non-trivial contributions to the VLA literature ¹⁸¹⁹²⁰. The architectural choices — separating visual geometry from language priors, decoupling visual and motor streams, operating at 36 Hz for real-time control — address real and documented limitations of prior VLA approaches. The two confirmed deployments (Wuhan apartments, FAW factory) are real, even if modest in scale ⁷⁹.

The funding is real and the investor quality is notable. Huawei Hubble does not invest in companies without technical due diligence ⁵. Fortune Capital and Lion City Capital are credible institutional investors ¹³. The oversubscription of the B2 round suggests genuine investor demand, not a manufactured narrative ¹⁰.

What Is Claimed But Unverified

The benchmark rankings — GigaBrain-0 at #1 on RoboChallenge, GigaWorld-1 at #1 on WorldArena, GigaWorld-Policy at #1 on RoboCasa365 — are vendor-reported figures disseminated through trade press ¹³. The independence of these benchmark platforms from the company is not established in the public record. RoboChallenge, WorldArena, and RoboCasa365 are referenced but not described in sufficient detail to assess their methodology, task diversity, or evaluation rigour. The specific claim of outperforming pi0.5 by approximately 10 percentage points is precise enough to be falsifiable, but no independent replication is present in the dossier.

The "85% average success rate on grasping, assembly, and sorting" figure ¹ is similarly unverified. Success rate in manipulation benchmarks is highly sensitive to task definition, object set, lighting conditions, and whether the evaluator is the same team that trained the model. Without a published evaluation protocol and independent replication, this number is a marketing figure.

The claim that GigaWorld-0 is "the world's first systematically introduced embodied world model" ⁵ is a superlative that is almost certainly false in the strict sense — world models for robotics have been explored in academic literature for several years — and is at best a claim about a particular systematic framing or product packaging.

What Is Demonstrably Overstated

The vendor language around deployment scale is the clearest case of overstatement. "Large-scale mass production and implementation" for the DriveDreamer series ⁵ is contradicted by the actual deployment evidence: 100 units in scenario testing in Wuhan, one factory pilot, and a three-year forward plan. These are pilot-stage activities, not mass production. The gap between the language and the evidence is not a minor rhetorical flourish; it is a material misrepresentation of commercial maturity.

The "30+ automaker and AV company partners" claim ⁵ is unverified and the term "partner" is undefined. In the Chinese technology ecosystem, this figure could include companies that have signed non-binding memoranda of understanding, attended a product demonstration, or participated in a data-sharing pilot. Without named customers confirming active, paid engagements, the figure is not analytically useful.

The Ugly: Structural Risks

Three structural risks deserve explicit statement.

First, the data flywheel is the company's stated competitive moat, but the target of 1 million hours of training data by end of 2026 ⁹ is an extraordinary ambition. At 36 Hz operation, 1 million hours of robot data would require approximately 114 years of continuous single-robot operation, or proportionally fewer robots operating in parallel. The company's Maker M01, U-01, and E-01 data collection hardware suggest a multi-source strategy including human demonstration and synthetic generation, but the quality and diversity of synthetic data generated by an unvalidated world model is itself uncertain. There is a circularity risk: using GigaWorld to generate training data for GigaBrain, which is then evaluated on benchmarks that may themselves be influenced by the same data distribution.

Second, the capital intensity of the current trajectory is high. 3.5 billion yuan in three months is a rapid burn rate for a company with no disclosed revenue. The 1,000-robot Longsheng agreement and the Wuhan programme are not yet revenue-generating at scale. If the next funding round is delayed or the benchmark claims are challenged, the company's runway becomes a critical variable.

Third, the residential data collection context raises questions that the company has not publicly addressed. A robot operating in a private apartment collects environmental data — spatial maps, object locations, resident behaviour patterns — that is sensitive under any reasonable privacy framework. The absence of any public statement on data governance, consent mechanisms, or regulatory compliance is a gap that will need to be filled before any serious institutional deployment outside China.

Claim tracker

GigaBrain-0 ranks #1 on RoboChallenge with a 51.67% task success rate, approximately 10 percentage points ahead of π0.5, and GigaWorld-Policy ranks #1 on RoboCasa365, beating Nvidia GR00T N1.5 and π0.5.Not supported

All benchmark rankings are self-reported or relayed through company press releases in trade media [1][7][11]; the independence of the benchmark platforms (RoboChallenge, RoboCasa365, WorldArena) from Jijia Vision is not established in any supplied source, and no third-party audit or neutral leaderboard administrator is cited.

100 SeeLight S1 units have been deployed in Wuhan Optics Valley talent apartments from May 31, 2026 for real-world scenario testing, with a broader seed-user program planned for 2027.Unknown

Multiple news sources corroborate the 100-unit Wuhan deployment date and location [7][9][10], but no independent resident feedback, property management confirmation, or journalist on-site report verifies actual robot operation or task performance in the apartments.

The TriVLA model operates at ~36 Hz with a triple-system architecture (VLM + dynamics perception + motor policy), enabling real-time feedback and step-wise error correction for general robot control.Unknown

The TriVLA architecture and 36 Hz operating frequency are described in an arXiv preprint [20] affiliated with Sun Yat-sen University and X-Era AI Lab, providing technical detail, but the paper is not peer-reviewed and no independent replication or real-robot benchmark by a third party is present in the supplied facts.

Jijia Vision raised approximately 3.5 billion yuan across multiple rounds within ~3 months in 2026, with investors including Huawei Hubble, Fortune Capital, and Lion City Capital.Unknown

The 3.5 billion yuan total and named investors are corroborated across multiple trade news outlets [1][3][5][6][10][16], but no regulatory filing, stock exchange disclosure, or independent financial audit confirms the figures; valuation was explicitly not disclosed, and round details rely on company announcements relayed by media.

---

12Future Scenarios

The following scenarios are editorial constructions based on the available evidence. They are not forecasts; they are structured possibilities intended to frame monitoring priorities. Each is assigned a rough plausibility assessment based on the current evidence base.

Scenario A: Benchmark Validation and Platform Breakout (Plausible, ~25%)

Independent researchers or a credible third-party evaluation body replicate Jijia Vision's benchmark claims within the next 12 months. GigaBrain-0's performance on RoboChallenge and RoboCasa365 is confirmed to be competitive with or superior to pi0.5 and GR00T N1.5. This validation triggers a wave of enterprise interest: the Longsheng 1,000-robot agreement accelerates, additional automotive manufacturers sign deployment contracts, and the Wuhan residential programme scales to multiple cities. The company raises a Series C at a disclosed valuation above 10 billion yuan and begins preparations for a domestic A-share or Hong Kong IPO.

In this scenario, Jijia Vision becomes a genuine platform company — the world-model stack licenced to hardware OEMs, the GigaBrain model offered as a cloud API, and the SeeLight S1 as a reference hardware design. The data flywheel compounds: more deployments generate more data, which improves the model, which enables more deployments.

This scenario requires the benchmark claims to be substantially accurate, the world-model architecture to generalise beyond the training distribution, and the company to execute on manufacturing scale-up without the quality and supply-chain problems that have plagued other Chinese hardware startups.

Scenario B: Niche Industrial Deployment with Slow Consumer Progress (Most Plausible, ~40%)

The FAW factory deployment proves sufficiently productive to attract two or three additional automotive or electronics manufacturing customers by end of 2026. The Longsheng agreement proceeds but at a slower pace than the three-year plan implies — perhaps 150-200 units deployed by end of year two. The residential programme remains in pilot mode through 2027, with the elder-care application delayed by regulatory and liability concerns.

The benchmark claims are neither definitively validated nor definitively refuted; they exist in the ambiguous space of unverified vendor assertions that the trade press continues to report uncritically. The company's valuation stabilises rather than compounds, and the next funding round takes longer than three months to close.

In this scenario, Jijia Vision is a viable but not dominant player: a credible industrial manipulation software company with a consumer hardware aspiration that remains perpetually 18 months from realisation. This is the trajectory of many Chinese robotics startups of the 2018-2022 vintage.

Scenario C: Benchmark Deflation and Competitive Squeeze (Plausible, ~25%)

A credible independent evaluation — perhaps from a university lab, a competing company's technical team, or a journalist with access to the evaluation environment — finds that the RoboChallenge and RoboCasa365 results are not reproducible under standardised conditions, or that the task set was narrow enough to be unrepresentative of general manipulation capability. Simultaneously, Physical Intelligence or Nvidia releases a model update that clearly outperforms GigaBrain-0 on the same benchmarks.

The narrative damage is significant. Investors who backed the B2 round on the strength of the benchmark claims reassess. The Longsheng agreement is renegotiated or delayed. The Wuhan residential programme is quietly wound down after the scenario-testing phase without a commercial follow-on. The company pivots to a narrower industrial niche — perhaps automotive mould handling specifically — where it can demonstrate genuine value without the burden of general-purpose claims.

This scenario does not imply fraud or failure; it implies a company that over-claimed on generality and is forced to find a more defensible market position. Many successful industrial robotics companies occupy exactly this kind of focused niche.

Scenario D: Capital Exhaustion and Consolidation (Lower Plausibility, ~10%)

The combination of high burn rate, unverified benchmarks, and a tightening Chinese venture market results in a failed Series C. The company's runway — undisclosed but presumably substantial given the 3.5 billion yuan raised — is consumed by compute costs, hardware manufacturing, and the data collection programme without generating sufficient revenue to demonstrate a path to profitability. A larger player — Huawei, Alibaba, or a state-owned enterprise — acquires the technology stack and team at a distressed valuation.

This scenario is lower probability given the quality of investors and the current Chinese policy environment favouring physical AI investment, but it is not negligible. The history of Chinese AI hardware startups includes several well-funded companies that failed to bridge the gap between impressive demos and profitable deployment.

The Autonomous Driving Wild Card

A scenario not captured above: the world-model work that underpins GigaWorld proves more valuable for autonomous driving simulation than for embodied robotics, and a major AV company (SAIC, BYD, or a tier-one supplier) acquires or exclusively licences the DriveDreamer stack. This would represent a pivot from the current embodied-robotics narrative but could be commercially superior. The company's stated 30-plus AV company relationships, if even partially real, provide the relationship infrastructure for this outcome.

---

13What to Watch: A Live Monitoring Checklist

The following indicators are the most diagnostically useful signals for assessing Jijia Vision's trajectory over the next 12-24 months. They are ordered by analytical priority.

Tier 1: Decisive Indicators (High Priority)

Independent benchmark replication. The single most important event to watch is whether any entity outside Jijia Vision's control replicates the RoboChallenge, WorldArena, or RoboCasa365 results. This could come from a university lab, a competing company's technical publication, or a structured evaluation by a robotics benchmark organisation. Replication confirms the technical claims; failure to replicate or absence of replication after 12 months is itself a signal.

Named customer confirmation. A public statement from a named customer — not a Jijia Vision press release, but a statement from the customer's own communications — confirming active, paid deployment of GigaBrain or SeeLight S1 in a production environment. The FAW deployment is the closest current example, but the public record does not include FAW's own confirmation of the deployment's scope or performance.

Wuhan residential programme outcome. The scenario-testing phase in Wuhan Optics Valley apartments was scheduled to begin 31 May 2026 ⁷. By Q4 2026, there should be sufficient operational data to assess whether the programme is proceeding to the planned 2027 seed-user expansion or has been quietly wound down. Watch for: resident feedback (positive or negative), local media coverage, and any municipal government statements about the programme.

Peer-reviewed publication of GigaBrain or GigaWorld. The current research output (VGA, FutureVLA, TriVLA) is on arXiv but not yet peer-reviewed in the dossier ¹⁸¹⁹²⁰. Acceptance at a top-tier venue (CoRL, ICRA, NeurIPS, ICLR) would provide independent technical validation of the underlying methodology, even if it does not directly validate the product benchmarks.

Tier 2: Important but Contextual Indicators (Medium Priority)

Longsheng Technology deployment progress. The 1,000-robot, three-year agreement is a significant commercial commitment ³. Quarterly or annual updates on actual units deployed versus the plan will reveal whether the agreement is proceeding or stalling. Watch for: Longsheng's own communications, Wuxi municipal government announcements, and any supply-chain disclosures.

Compute infrastructure disclosure. Whether Jijia Vision is training on Huawei Ascend, Nvidia (if accessible), or a hybrid infrastructure has significant implications for benchmark comparability and for the company's exposure to semiconductor export controls. Any disclosure — even indirect, through job postings or infrastructure partnership announcements — is analytically useful.

Data collection programme scale. The 1 million hours target by end of 2026 ⁹ is ambitious. Watch for: announcements of new data collection partnerships, expansion of the Maker M01 / U-01 / E-01 hardware programme, and any disclosure of actual data volume accumulated.

Regulatory filings and safety certifications. For the residential deployment to scale beyond scenario testing, the SeeLight S1 will need to satisfy Chinese product safety standards (GB standards for service robots) and potentially obtain municipal or provincial approval for residential operation. Watch for any certification announcements.

Additional investor or strategic partner announcements. The B2 round closed in June 2026 ¹⁰. A Series C announcement — its timing, size, and investor composition — will signal whether the capital market continues to validate the company's trajectory.

Tier 3: Background Monitoring (Lower Priority but Useful)

Talent movements. Key hires or departures from the research team, particularly researchers associated with the VGA, FutureVLA, or TriVLA papers, are leading indicators of organisational health.

International expansion signals. Any announcement of deployment, partnership, or regulatory engagement outside China would represent a significant step change in the company's ambition and risk profile.

Competitor benchmark updates. If Physical Intelligence, Nvidia, or another competitor publishes updated benchmark results on RoboChallenge or RoboCasa365, the relative positioning of GigaBrain-0 will become clearer — either confirming or undermining the current claims.

Chinese government policy signals. Changes to China's robotics industry support policies, data governance regulations for residential AI systems, or export-control frameworks affecting AI model weights could materially affect the company's operating environment.

Media coverage tone shift. The current media coverage is uniformly positive and largely based on company press releases. A shift toward critical or investigative coverage — in Chinese-language tech media such as 36Kr, LatePost, or Caixin — would be an early warning signal worth tracking.

---

14Sources and Methodology

Source List

¹ "Jijia Vision", an Embodied Intelligence Company, Raised 500M Yuan in 3 Months, Completed 200M Yuan Series A2 Financing and Launched Native Model and Ontology of Physical AGI — https://eu.36kr.com/en/p/3586459879750789

² Computer Vision Technology Costs: Key Factors & Use Cases - it-jim — https://www.it-jim.com/blog/computer-vision-costs (Not used: misattributed source, excluded from analysis)

³ Seeds | Jijia Vision Secures Another 1 Billion Yuan in Financing, Raising 3.5 Billion in Three Months | Gasgoo — https://autonews.gasgoo.com/articles/news/seeds-jijia-vision-secures-another-1-billion-yuan-in-financing-raising-35-billion-in-three-months-2066465297387855873

⁴ Pricing | Cloud Vision API - Google Cloud — https://cloud.google.com/vision/pricing (Not used: misattributed source, excluded from analysis)

⁵ Jijia Vision Secures Hundreds of Millions in Series A1 Strategic Financing Co-led by Huawei Hubble and Huakong Fund: Pioneering Physical AI's Ultimate Technical Route with "World Model" — https://eu.36kr.com/en/p/3536726905445508

⁶ Seeds | Jijia Vision Secures Another 1 Billion Yuan in Financing, Raising 3.5 Billion in Three Months | Gasgoo (ICV edition) — https://autonews.gasgoo.com/articles/icv/seeds-jijia-vision-secures-another-1-billion-yuan-in-financing-raising-35-billion-in-three-months-2066465297387855873

⁷ GigaAI SeeLight S1: What Wuhan Trial Proves — ui44 Blog — https://ui44.com/blog/gigaai-seelight-s1-home-robot-wuhan-trial

⁸ 具身通用机器人极佳视界完成10亿元B2轮融资：AI大模型赋能，瞄准养老陪护与工业协作双场景 | 聚看JUCAP — https://jucap.com/news/tech/20260616-08-jijia-vision-1b-funding/

⁹ 极佳视界完成10亿元Pre-B轮融资，「世界模型」驱动通用机器人加速进入千行百业 - 36氪 — https://www.36kr.com/p/3710076436557954

¹⁰ 极佳视界完成10亿元B2轮融资，3个月内累计融资35亿元|机器人|评测_网易科技 — https://www.163.com/tech/article/KVF97C0R00097U7T.html

¹¹ Jijia Shijie AI Beats Google, Secures ¥1B Funding — https://www.edgen.tech/news/post/jijia-shijie-ai-beats-google-secures-1b-funding

¹² 极佳视界再获10亿融资 黄冠称物理AGI将直接作用于真实物理世界|本体|机器人|agi_网易订阅 — https://www.163.com/dy/article/KVFVGQR20511A0EF.html

¹³ 极佳视界再获全球顶配阵容10亿投资：加速物理AGI突破_腾讯新闻 — https://news.qq