Galbot (Beijing Galaxy General Robot Co., Ltd.)

The best-funded humanoid robotics startup you have probably never heard of — and why that matters

---

How to Read This Report

This report separates four categories of claim throughout. Readers should weight them accordingly.

Inline citations use bracketed numerals keyed to the Sources list in §14. Sources are drawn exclusively from the supplied research dossier. Where the dossier is thin, this report says so rather than padding with inference dressed as fact.

---

01Executive Overview

Galbot — the trading name of Beijing Galaxy General Robot Co., Ltd. — is, by any conventional financial measure, the most consequential humanoid robotics company most Western industry observers have not yet examined closely. Founded in May 2023 by Wang He, a researcher from Peking University, the company has raised approximately $800 million to $930 million in total across multiple rounds in under three years, carries a valuation of roughly $3 billion, and is described across multiple independent news sources as the highest-valued unlisted humanoid robotics firm in China ⁷¹¹¹⁵. It is preparing for a listing on the Hong Kong Stock Exchange ⁷¹¹.

The flagship product is the G1: a semi-humanoid mobile manipulator standing 173 cm tall, riding an omnidirectional wheeled base, equipped with two arms spanning 190 cm, and driven by a proprietary AI stack that Galbot calls AstraBrain ¹³. A heavier-duty industrial variant, the S1, carries a 50 kg dual-arm payload and features autonomous battery swapping ⁸¹²²⁵. Neither platform walks on legs; the deliberate choice of a wheeled base is an engineering trade-off that sacrifices stair-climbing for stability, payload capacity, and operational reliability in the flat-floor environments that constitute the overwhelming majority of industrial and retail floor space.

Commercial deployments are confirmed across multiple independent news sources: approximately ten Beijing pharmacies and retail stores, warehouse logistics operations, and — most significantly — CATL's battery production line in Luoyang, Henan Province ⁶⁸⁹. Orders from CATL, Bosch, Toyota, BAIC Group, and SAIC Motor are reported by Technode citing a company statement, placing the claimed order book in the several-thousand-unit range ¹¹. These figures are COMPANY CLAIMS corroborated by named-customer involvement but not independently audited.

The technology stack is unusually research-dense for a company of this age. Peer-reviewed publications from affiliated institutions — Tsinghua University, Peking University, BAAI, CASIA, and others — validate specific sub-capabilities: a handover task framework trained entirely on synthetic data ¹⁹, a lifelong reinforcement-learning navigation system ²¹, and a 1-billion-parameter latent dynamics action model ²². These papers do not prove 24/7 autonomous commercial operation at scale, but they do establish that the underlying research programme is substantive and not merely marketing theatre.

The central tension in any analysis of Galbot is the gap between what the research validates and what the company claims commercially. Vendor assertions of "100% autonomous performance without teleoperation" and "24/7 autonomous operation for over a year" are consistent across sales and media materials but have not been independently audited. The research corroborates the plausibility of autonomous sub-task execution; it does not confirm uninterrupted commercial autonomy at the level claimed. That distinction matters enormously for any procurement or investment decision.

EDITORIAL INFERENCE: Galbot is a serious industrial robotics company with genuine technical depth, substantial state-backed capital, and real commercial deployments. It is not a paper company. It is also not yet a proven at-scale autonomous industrial operator in the sense that, say, a decade of Kiva/Amazon Robotics deployment data would establish. The honest position is that it sits somewhere between those poles — and the next 18 months of IPO preparation will force a degree of operational transparency that the company has not yet had to provide.

Latest news

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

---

02The Galbot Story

Founding and Provenance

Beijing Galaxy General Robot Co., Ltd. was incorporated in May 2023 ¹¹¹. Its founder, Wang He, came from Peking University, one of China's two most prestigious research universities and a consistent source of robotics and AI talent ¹⁰. The company's research collaboration network — which spans Tsinghua University, Peking University, BAAI (Beijing Academy of Artificial Intelligence), CASIA (Chinese Academy of Sciences Institute of Automation), USTC, Shanghai AI Lab, Shanghai Qi Zhi Institute, Sun Yat-sen University, Zhejiang University, the University of Adelaide, and NVIDIA — reads less like a startup's advisory board and more like a coordinated national research programme ^19202122. EDITORIAL INFERENCE: This network density is not accidental. It reflects both Wang He's academic connections and the broader Chinese state strategy of fusing university research with commercially oriented robotics ventures.

The Funding Trajectory

The company's fundraising history is remarkable in its speed and scale. The dossier does not provide a complete round-by-round breakdown with precise dates for every tranche, but the convergent picture from multiple independent news sources is as follows.

The March 2026 round is the most significant not for its size alone but for its composition. The National AI Industry Investment Fund is a direct instrument of Chinese state industrial policy. Sinopec is one of the world's largest energy and chemicals companies. The China Integrated Circuit Industry Investment Fund — colloquially known as the "Big Fund" — is the vehicle through which Beijing channels capital into strategic semiconductor and advanced manufacturing sectors ¹⁰. CATL's earlier participation as a named investor in the mid-2025 round ¹⁶ means China's dominant battery manufacturer is simultaneously a customer, a technology partner (supplying batteries for the S1 ⁸¹²), and an equity holder. This is not a conventional venture capital structure; it is a strategic industrial ecosystem.

The Gasgoo report citing "nearly 5 billion yuan raised in three months" ¹⁵ and the cnmra.com report referencing "6 billion yuan in funding" ¹⁷ suggest either that additional tranches were announced in rapid succession or that different sources are counting cumulative totals differently. The dossier flags a confidence level of 0.92 on the $800M–$930M total, with the upper bound reflecting inclusion of the latest state-backed round. EDITORIAL INFERENCE: The precise total is less important than the structural fact: Galbot has more capital than any Western humanoid robotics startup of comparable age, and a significant portion of that capital comes from entities whose investment decisions are shaped by national industrial strategy rather than purely commercial return expectations.

The Name and the Brand

The company operates under two registered names: Beijing Galaxy General Robot Co., Ltd. and Beijing Galbot AI Co., Ltd. ¹. The "Galbot" brand — a contraction that evokes both "galaxy" and "robot" — is used consistently in international marketing. The G1 product name echoes a naming convention common in Chinese robotics (compare Unitree's G1), though the platforms are architecturally distinct.

Strategic Positioning from Day One

What distinguishes Galbot's founding narrative from many contemporaries is the explicit decision to target industrial deployment from the outset rather than building a research demonstrator and pivoting to commercialisation later. The G1's wheeled base, its 15 kg single-arm payload, its 5,000-SKU goods-handling claim, and its 1-day store deployment claim ³⁹ are all consistent with a product designed for immediate industrial utility rather than for research prestige. The S1's 50 kg dual-arm payload and autonomous battery swapping ⁸²⁵ extend this logic into heavy manufacturing. Whether the execution matches the intent is the subject of later sections of this report; the founding intent itself is legible and coherent.

---

03Product Portfolio: What Galbot Actually Sells

The G1: Core Platform

The Galbot G1 is the company's primary commercial product. It is a semi-humanoid mobile manipulator: humanoid from the waist up (two arms, a torso, a sensor head), wheeled from the waist down (an omnidirectional base). This architecture is a deliberate departure from fully bipedal humanoids. The trade-off is explicit: no stair-climbing capability in exchange for superior payload stability, lower mechanical complexity in the locomotion system, and higher operational reliability on flat industrial and retail floors ¹²³.

Verified physical specifications (convergent across multiple commerce sources, confidence 0.88 ¹²³⁴):

The battery life conflict deserves specific attention. A commerce/reseller listing claims "10+ hours battery life" ¹, which aligns with Galbot's marketing narrative. The humanoid.guide aggregator states 2 hours ², with an explicit caveat that its data is unverified. Neither figure derives from an independent teardown or controlled operational test. The 10+ hour figure may refer to a specific configuration, duty cycle, or operational mode; the 2-hour figure may reflect a different variant or a conservative real-world estimate. UNKNOWN: No independent battery life test data exists in the public domain at the time of this report. Buyers should treat both figures with caution and request vendor-supplied test data under representative duty cycles.

Compute Variants

The G1 is offered in at least two compute configurations ¹³⁴:

• G1 Standard: Equipped with Galbot's proprietary AstraBrain AI system, described as a "brain-cerebellum-neural control architecture." This is the baseline commercial configuration.
• G1 Premium: Equipped with an NVIDIA Jetson Thor compute module, positioned at a higher price point. The Jetson Thor is NVIDIA's robotics-specific SoC, designed for transformer-based inference workloads at the edge ⁵.

EDITORIAL INFERENCE: The dual-compute strategy is commercially sensible. The proprietary AstraBrain path preserves margin and supply chain control; the Jetson Thor path offers customers a known, supported compute platform with an established software ecosystem and the credibility of the NVIDIA brand. The Premium variant also signals that Galbot's AI stack can run on third-party hardware, which reduces customer lock-in risk perception.

The S1: Heavy Industrial Variant

The S1 is positioned as a heavy-duty industrial platform, differentiated from the G1 primarily by payload capacity and operational endurance features ⁸¹²²⁵:

The autonomous battery swapping capability is operationally significant. It means the S1 can, in principle, operate continuously across shifts without human intervention for recharging — a genuine differentiator for 24/7 manufacturing environments. COMPANY CLAIM: Galbot states the S1 achieves this in practice at the CATL Luoyang facility. UNKNOWN: No independent operational data on swapping cycle times, failure rates, or uptime percentages has been published.

The CATL partnership on the S1 is structurally notable: CATL supplies the batteries, is an equity investor, and is a named deployment customer ⁶⁸¹²¹⁶. This vertical integration of the supply chain through an equity relationship is unusual and creates both advantages (preferential battery access, co-development) and risks (dependency on a single strategic partner for a critical component).

Pricing

Pricing is one of the more contested areas in the dossier.

The $55,000 figure is self-labeled as unverified by its source and should be disregarded for any analytical purpose ². The spread between the Chinese domestic price (~$87,000–$97,000) and the international listing ($119,995) is consistent with import/distribution markup and is not inherently suspicious. EDITORIAL INFERENCE: At $87,000–$97,000 in the domestic market, the G1 is priced above Unitree's H1 and G1 platforms but below Boston Dynamics' Spot ecosystem for comparable industrial utility. Whether that price point is justified by operational performance is a question the market is currently answering through the deployment programmes described in §7.

Software and AI Stack

The AI stack is described in more detail in §4, but from a product perspective the key commercial claims are:

• Handles 5,000+ types of goods (COMPANY CLAIM) ³⁹
• 95–97% grasp success rate (COMPANY CLAIM) ³
• 1-day deployment into new retail stores (COMPANY CLAIM) ⁹
• Pure vision-based navigation without floor markers or infrastructure modification (COMPANY CLAIM, partially corroborated by research ²⁰²¹)

The "1-day deployment" claim is commercially important because it directly addresses the integration cost objection that has historically slowed industrial robot adoption. If accurate, it dramatically lowers the total cost of deployment relative to systems requiring weeks of environment mapping, marker installation, or custom integration work. EDITORIAL INFERENCE: The research on AllDayNav ²¹ and the embodied navigation foundation model ²⁰ provides a plausible technical basis for rapid deployment, but the 1-day figure itself is a vendor claim without independent verification.

Products & versions

Galbot G1

Semi-humanoid mobile manipulator with wheeled omnidirectional base, dual arms, 25 DOF, 173 cm tall, 15 kg single-arm payload, and AI-driven autonomy for industrial and retail deployments.

Galbot G1 Premium

Enhanced variant of the G1 powered by NVIDIA Jetson Thor, offering higher compute performance for more demanding autonomous task execution.

Galbot S1

Heavy-duty semi-humanoid variant with 50 kg dual-arm payload, vision-only centimeter-level positioning, 360° omnidirectional obstacle avoidance, CATL batteries, 8-hour battery life, and autonomous battery swapping.

---

04Technology Stack: Strengths and the Work That Remains

Architecture Overview

Galbot's AI and control architecture is described under the "AstraBrain" brand, which the company characterises as a "brain-cerebellum-neural control" hierarchy ¹³. This three-tier framing — high-level reasoning, mid-level motion planning, low-level motor control — is standard in modern robot systems and is not unique to Galbot. What is more distinctive is the specific set of models and frameworks the company has developed and, in several cases, published in peer-reviewed venues.

Navigation: AllDayNav and the Embodied Navigation Foundation Model

Two peer-reviewed publications address Galbot's navigation capabilities directly.

AllDayNav ²¹ is described as a lifelong navigation framework using real-world reinforcement learning. The key claim is that the system can navigate in changing environments without requiring periodic re-mapping or human recalibration — the "lifelong" property. The paper is affiliated with Galbot's research network and has been published on arXiv. VERIFIED FACT: The paper exists and describes real-world experiments. EDITORIAL INFERENCE: The existence of a peer-reviewed paper does not confirm commercial deployment performance, but it does establish that the navigation approach is grounded in a specific technical methodology that can be scrutinised and replicated.

The Embodied Navigation Foundation Model ²⁰ describes a cross-embodiment navigation model trained on 8 million samples. The scale of the training dataset is notable; 8 million samples represents a substantial investment in data collection and curation. VERIFIED FACT: The paper reports training on 8M samples and describes the model architecture. COMPANY CLAIM: Galbot states this enables pure vision-based navigation without floor markers. The research provides a plausible technical basis for this claim but does not constitute independent commercial validation.

Manipulation: GraspVLA, GroceryVLA, and the VLA Family

Galbot has developed at least three Vision-Language-Action (VLA) models: GraspVLA, GroceryVLA, and a third that is not named in the dossier ³. VLA models are the current frontier approach to generalised robot manipulation — they combine visual perception, language understanding, and action generation in a single learned policy. The grocery-specific model (GroceryVLA) is directly relevant to the retail pharmacy deployments described in §7.

The MobileH2R paper ¹⁹ addresses human-to-mobile-robot handover, a specific and practically important sub-task for retail and warehouse environments where robots must receive objects from human workers. The paper's key methodological claim is that the system was trained exclusively on synthetic data — no real-world demonstrations were required. VERIFIED FACT: The paper reports a +15% improvement over baselines in handover tasks in real-world testing. This is a peer-reviewed result, not a vendor claim. The absolute success rate of 95–97% cited in marketing materials ³ is a COMPANY CLAIM that is not reported in the research paper and cannot be independently verified from the dossier.

Foundation Model: LDA-1B

The LDA-1B paper ²² describes a 1-billion-parameter latent dynamics action model trained via "universal embodied data ingestion." The model is designed to work across multiple robot embodiments — a property that, if it generalises robustly, would allow Galbot to leverage training data from diverse robot platforms rather than being limited to data collected on its own hardware. VERIFIED FACT: The paper exists and describes the architecture. EDITORIAL INFERENCE: A 1B parameter model is modest by large language model standards but is at the larger end of what is currently practical for real-time robot control inference on edge hardware. The Jetson Thor in the G1 Premium is specifically designed for this class of workload ⁵.

Training Methodology: Sim2Real and the EI-30k Dataset

VERIFIED FACT (from peer-reviewed publications): Galbot's training methodology relies heavily on sim-to-real transfer and synthetic data generation, with the explicit goal of avoiding dependence on large-scale real-world demonstration collection ¹⁹²². The EI-30k dataset — described as 30,000+ hours of human and robot trajectories — is used for imitation learning and reinforcement learning ³. EDITORIAL INFERENCE: The synthetic data approach is strategically important because it allows the company to scale training data without the labour-intensive and expensive process of teleoperated demonstration collection. If the sim-to-real transfer is robust, this is a genuine competitive advantage. If the transfer gap is large, it is a source of brittleness in real-world deployment.

Strengths

1. Research depth: Multiple peer-reviewed publications from a credible multi-institution network validate specific sub-capabilities. This is not common for a company less than three years old.
2. Synthetic data training: The ability to train on synthetic data at scale, validated in the MobileH2R paper ¹⁹, addresses a key bottleneck in robot learning.
3. Navigation without infrastructure: Pure vision-based navigation, if it performs as described, eliminates a major deployment friction point.
4. Vertical integration of compute: The dual AstraBrain/Jetson Thor strategy provides flexibility without abandoning proprietary development.
5. CATL battery partnership: Access to world-class battery technology for the S1 platform is a genuine hardware advantage.

The Work That Remains

1. Battery life ambiguity: The unresolved conflict between 2-hour and 10+ hour battery life figures ¹² is not a minor discrepancy. For 24/7 industrial operation, battery endurance is a critical operational parameter. The absence of independent test data is a gap.
2. Stair-climbing and unstructured environments: The wheeled base is a deliberate trade-off, but it limits the addressable environment to flat floors. As industrial deployments expand to facilities with level changes, ramps, or outdoor connections, this constraint will matter.
3. Absolute success rate validation: The 95–97% grasp success rate ³ is a vendor claim without peer-reviewed corroboration. At scale, the difference between 95% and 97% success on millions of picks per year is operationally significant.
4. Long-horizon task reliability: Peer-reviewed papers validate specific sub-tasks (handover, navigation, grasping). No published research addresses multi-step, long-horizon task completion in uncontrolled commercial environments over extended periods.
5. Failure mode transparency: No public data exists on failure modes, error recovery behaviour, or maintenance intervals. These are standard requirements for industrial procurement decisions.
6. Dual-arm payload discrepancy: The G1's dual-arm payload of 5 kg is substantially lower than its single-arm payload of 15 kg ¹³. This is physically unusual — dual-arm payloads are typically higher than single-arm, not lower. UNKNOWN: Whether this reflects a genuine mechanical constraint, a measurement methodology difference, or a data error in the source material is not clear from the dossier.

---

05Research, Papers, Authors and Labs

Research Programme Overview

Galbot's research output is unusually substantial for a company of its age and commercial focus. The peer-reviewed publications identified in the dossier span navigation, manipulation, foundation model development, and training methodology — covering the full stack of capabilities required for autonomous mobile manipulation. All identified papers are available on arXiv, which is standard practice in the robotics and AI research community.

Key Publications

MobileH2R: Learning Generalizable Human to Mobile Robot Handover Exclusively from Scalable and Diverse Synthetic Data ¹⁹

This paper addresses the human-to-robot handover problem for mobile platforms — specifically, how a wheeled robot can receive objects from a human in a generalised, unstructured way. The methodological contribution is training entirely on synthetic data, which the authors validate with real-world experiments showing a +15% improvement over baselines. The institutional affiliations listed on the paper include members of Galbot's known research network. This is the strongest piece of independent technical validation in the dossier for Galbot's manipulation capabilities.

Embodied Navigation Foundation Model ²⁰

This paper describes a cross-embodiment navigation model trained on 8 million samples. The scale of the training corpus and the cross-embodiment design are the key contributions. The paper is relevant to Galbot's claim of rapid deployment into new environments without infrastructure modification.

AllDayNav: Lifelong Navigation via Real-World Reinforcement Learning ²¹

This paper is directly relevant to Galbot's commercial claim of 24/7 autonomous operation. Lifelong navigation — the ability to continue operating reliably as the environment changes over time — is a prerequisite for uninterrupted commercial deployment. The use of real-world reinforcement learning (rather than purely simulation-based training) is notable; it implies the system has been tested in real environments over extended periods. EDITORIAL INFERENCE: This paper provides the strongest technical basis for Galbot's navigation reliability claims, though the gap between a research paper's controlled experimental conditions and a commercial deployment's full operational complexity remains.

LDA-1B: Scaling Latent Dynamics Action Model via Universal Embodied Data Ingestion ²²

This paper describes the 1B-parameter foundation model that underpins Galbot's action generation. The "universal embodied data ingestion" framing suggests the model is designed to learn from heterogeneous data sources across robot types, which is a current research frontier. The paper's existence confirms that Galbot is investing in foundation model research at a scale comparable to well-resourced academic labs.

Institutional Network

VERIFIED FACT: The following institutions appear as co-authors or affiliates on Galbot-associated research papers ^19202122:

• Tsinghua University
• Peking University
• BAAI (Beijing Academy of Artificial Intelligence)
• CASIA (Chinese Academy of Sciences Institute of Automation)
• USTC (University of Science and Technology of China)
• Shanghai AI Lab
• Shanghai Qi Zhi Institute
• Sun Yat-sen University
• Zhejiang University
• University of Adelaide
• NVIDIA

The breadth of this network — spanning Beijing, Shanghai, Guangzhou, Hangzhou, Adelaide, and Santa Clara — is striking. EDITORIAL INFERENCE: This is not a single university spin-out with a narrow research base. It is a coordinated multi-institution programme that reflects both Wang He's academic network and the Chinese state's interest in accelerating embodied AI research through institutional collaboration.

What the Research Does and Does Not Prove

It is important to be precise about what peer-reviewed publication establishes and what it does not.

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

• EI-30k

  A 30,000+ hour dataset of human and robot trajectories used for imitation learning and sim-to-real transfer training of Galbot's manipulation models.

---

06Media Evidence Library: What the Videos Prove

The Video Corpus

The dossier contains six video sources associated with Galbot ^52324252627. One additional video ²⁸ is clearly unrelated (an Enabot EBO Max review) and is disregarded. The six relevant videos span product demonstrations, investor/media coverage, and an international exhibition appearance. Each is assessed below against the editorial standard: a choreographed demo video is not proof of autonomous work.

Video-by-Video Assessment

[23] "This May Be China's Most Important Humanoid Robot Startup"

This appears to be a media or investor-oriented overview video. Based on its title and context in the dossier, it likely presents Galbot's narrative, funding story, and product demonstrations in a favourable light. EDITORIAL INFERENCE: This type of video is useful for understanding how the company presents itself and what capabilities it chooses to highlight, but it is not evidence of autonomous commercial operation. It is marketing content.

[24] "Galbot Robot in Action — IREX Japan 2025"

IREX (International Robot Exhibition) is a credible third-party venue. A robot performing tasks at IREX is operating in a public, observed environment — not a controlled lab. However, exhibition demonstrations are typically rehearsed, the environment is prepared, and the tasks are selected to showcase strengths. EDITORIAL INFERENCE: This video provides moderate evidence that the G1 can perform its demonstrated tasks in a semi-public environment. It does not prove autonomous commercial operation. The significance is that Galbot chose to exhibit internationally, signalling confidence in the platform's public performance.

[25] "Galbot S1: Heavy-Duty Robot"

This is a product demonstration video for the S1 variant. It likely shows the S1's payload capacity, autonomous battery swapping, and industrial manipulation tasks. EDITORIAL INFERENCE: Demonstration videos for specific capabilities (battery swapping, heavy lifting) are more informative than general "robot doing things" videos because the specific capability being demonstrated is clearly defined. If the S1 is shown autonomously swapping its own battery, that is meaningful evidence of that specific capability — though it does not prove the system performs this reliably over thousands of cycles.

[26] "Galbot G1: Our First Embodied AI Product Embodying the Future of Industrial and Home Autopilot"

This is a product launch or overview video. The title's reference to "home autopilot" alongside "industrial" is notable — it suggests Galbot is positioning the G1 for eventual consumer/home markets as well as industrial ones. EDITORIAL INFERENCE: This is primarily marketing content. The "home autopilot" framing should be treated with caution; the current commercial deployments are entirely industrial and retail, and home deployment raises substantially different regulatory, safety, and liability considerations.

[27] "Your humanoid tennis player is here!"

This video appears to show the G1 or a related platform performing tennis-related motions. EDITORIAL INFERENCE: Athletic demonstrations are a common technique in humanoid robotics marketing (compare Boston Dynamics' dancing videos, Agility Robotics' parkour content). They demonstrate dynamic motion capability and generate media attention. They do not demonstrate industrial utility. The tennis video should be read as a capability showcase and brand-building exercise, not as evidence of commercial readiness.

[5] "Galbot G1 Premium: The Swiftest Humanoid Robot Worker, Powered by NVIDIA Jetson Thor"

This video specifically highlights the G1 Premium's NVIDIA Jetson Thor configuration. The NVIDIA co-branding is commercially significant — NVIDIA's involvement lends credibility and signals that the compute platform meets NVIDIA's technical partnership standards. EDITORIAL INFERENCE: This video is more technically informative than general demonstrations because it addresses a specific hardware configuration. The "swiftest humanoid robot worker" claim in the title is marketing language and should not be taken as a verified benchmark.

What the Video Corpus Collectively Establishes

The video corpus is consistent with a company that has a functional, demonstrable product. It does not constitute evidence of the commercial performance claims that matter most to industrial buyers.

Media library

suara mobil polisi robot #shorts #funny #comedy #meme

YouTube

Galaxy General Robot officially releases the world's first humanoid robot general cerebellum GPT foundation model AstraBrain-WBC 0.5

Bilibili8.3k views

---

07Commercial Reality

What Is Actually Deployed

VERIFIED FACT (convergent across multiple independent news sources): Galbot has commercial deployments in the following categories ^6891113:

• Approximately 10 Beijing pharmacies and retail stores, with a stated target of 100 nationally ⁹
• Multiple warehouse logistics locations
• CATL's smart production line in Luoyang, Henan Province ⁶⁸
• IREX Japan 2025 exhibition (demonstration, not production deployment) ²⁴
• Hospitality and public venues (specific locations not disclosed)

The CATL deployment is the most significant and the most independently corroborated. Multiple independent news sources — CNEVPost ⁶, Interesting Engineering ⁸, Gasgoo ¹² — report on the CATL-Galbot collaboration, and CATL is a named equity investor ¹⁶, making the partnership relationship itself a verified fact. The operational details of the deployment (number of units, tasks performed, uptime, productivity metrics) are COMPANY CLAIMS that have not been independently audited.

The Order Book

COMPANY CLAIM (reported by Technode citing company statement ¹¹): Several thousand units have been ordered by CATL, Bosch, Toyota, BAIC Group, and SAIC Motor.

This claim requires careful parsing. "Orders" in the robotics industry can mean anything from binding purchase orders with delivery schedules to letters of intent, memoranda of understanding, or pilot programme commitments. The dossier does not specify which category these "orders" fall into. EDITORIAL INFERENCE: The involvement of CATL as both investor and customer, and the public nature of the CATL production line deployment, makes CATL's order the most credible. The Bosch and Toyota orders are plausible given those companies' active humanoid robotics programmes globally, but they are unverified by independent sources in the dossier. The BAIC and SAIC orders are consistent with the automotive sector's interest in humanoid robots for assembly tasks but are similarly unverified.

08Markets and Use Cases

Galbot's commercial strategy rests on a relatively focused set of industrial and semi-industrial verticals, each chosen to match the G1's physical architecture: a wheeled base that navigates flat or near-flat floors, two arms capable of dexterous manipulation, and a sensor suite oriented toward structured but dynamic environments. The company is not, at present, attempting to be all things to all buyers. That restraint is commercially sensible, even if the marketing language sometimes obscures it.

Industrial Manufacturing: The CATL Anchor

The highest-profile deployment is the partnership with CATL, the world's largest lithium-ion battery manufacturer, at its Luoyang, Henan Province facility ⁶⁸¹². The S1 variant — heavier, with a 50 kg dual-arm payload versus the G1's 15 kg single-arm figure — was developed in part to meet the physical demands of battery cell handling, where components are dense, fragile, and chemically sensitive ⁸²⁵. CATL's involvement is strategically significant beyond the revenue it represents: CATL is also an investor in Galbot ¹⁶, creating a customer-investor alignment that accelerates co-development but also raises questions about whether the deployment terms reflect arm's-length commercial reality.

The S1's autonomous battery swapping capability is directly relevant to factory floor operations, where downtime for recharging would interrupt production cadence ²⁵. An 8-hour stated battery life on the S1 ⁸ maps reasonably well to a single shift, though multi-shift continuous operation would require either multiple units or the swapping infrastructure to function reliably — neither of which has been independently verified at scale.

Beyond CATL, orders from Bosch and Toyota have been reported ¹¹, though the nature of those orders — pilot quantities, framework agreements, or firm production commitments — is not publicly disclosed. BAIC Group and SAIC Motor are also named ¹¹, suggesting the automotive supply chain is a primary target vertical. This makes structural sense: automotive assembly plants are large, flat, and already partially automated, reducing the navigation complexity the G1 would face in less structured environments.

Pharmaceutical Retail and Dispensing

Approximately ten Beijing pharmacies were operating G1 units as of the most recent reporting, with a stated target of 100 nationally ³²⁶. The pharmacy use case is instructive about where the G1 actually performs well: shelves are organised, SKUs are finite and well-labelled, the floor plan is stable, and the task repertoire — retrieve item, hand to customer or place in bag — is repetitive and bounded. These conditions are close to the structured environments in which the underlying VLA models were trained and validated.

The claim that the G1 handles 5,000+ types of goods ³ is a vendor figure and has not been independently tested across that full range. However, the pharmacy environment is one where even a narrower capability — say, reliable handling of 500–1,000 common SKUs — would be commercially useful. The 1-day deployment into new stores claim ³ is more difficult to evaluate: it likely refers to map acquisition and shelf-layout ingestion rather than full capability calibration, and the conditions under which it holds (store size, shelf density, lighting) are unspecified.

Warehouse Logistics and Fulfilment

Warehouse deployments are mentioned across multiple sources ³⁹¹³ but with less operational specificity than the pharmacy or CATL cases. The G1's omnidirectional wheeled base and pure vision-based navigation ²²⁰ are well-suited to warehouse aisles, and the AllDayNav lifelong reinforcement learning framework ²¹ is explicitly designed for environments that change over time — a critical requirement in active fulfilment centres where stock locations shift daily.

The claim of "24/7 autonomous operation for over a year in warehouses" ³ is the strongest autonomy assertion in the dossier and the least independently corroborated. It is plausible given the research depth behind the navigation stack, but no third-party logistics operator has publicly confirmed it. The absence of contradicting reports is not confirmation.

Hospitality and Public Venues

Galbot's appearance at IREX Japan 2025 ²⁴ and references to hospitality deployments ²⁶ suggest the company is exploring service-sector applications. These are harder markets: unstructured human traffic, unpredictable floor layouts, and a much wider range of interaction types. The G1's form factor — 173 cm, 85–88 kg, wheeled — is less threatening than a fully bipedal humanoid in a public space, but the interaction quality required for hospitality exceeds what is needed in a pharmacy or factory. No confirmed hospitality deployments with named operators appear in the dossier.

Use Case Suitability Matrix

The home-assistance reference in Galbot's own product video ²⁶ should be read as aspirational positioning rather than a near-term commercial target. At $87,000–$120,000 per unit and without stair-climbing capability, the G1 is not a consumer product by any realistic measure.

---

09Competitive Landscape

Galbot operates in a rapidly crowding field. The semi-humanoid mobile manipulator category — wheeled base, dual arms, AI-driven autonomy — has attracted serious capital and engineering talent globally, and Galbot's architectural choices place it in direct competition with several well-resourced rivals.

Direct Competitors: Semi-Humanoid Mobile Manipulators

Agility Robotics (Digit) is the most established Western competitor in the warehouse manipulation space, with a confirmed Amazon deployment programme. Digit is bipedal, which gives it stair-climbing capability the G1 lacks, but also makes it mechanically more complex and currently slower in flat-floor navigation. Agility was acquired by Hyundai in 2024, giving it manufacturing scale and automotive customer access that Galbot is still building toward.

1X Technologies (NEO) is pursuing a similar dual-arm, mobile-base architecture with a strong emphasis on learned behaviours from human demonstration data. It is Norway-headquartered with significant US presence and has raised substantial capital, though at a smaller scale than Galbot's reported $800M–$930M ⁷¹¹.

Apptronik (Apollo) targets manufacturing and logistics with a bipedal humanoid, backed by Google DeepMind collaboration and a GXO Logistics pilot. Its bipedal form factor is a differentiator for environments with stairs or uneven terrain.

Physically Intelligence (Pi) is less a hardware competitor and more a software-layer threat: if foundation models for robot manipulation become commoditised, Galbot's proprietary VLA stack loses some of its moat.

Chinese Domestic Competitors

The domestic competitive pressure on Galbot is intense and accelerating. China's humanoid robotics sector has attracted state policy support, and several well-capitalised startups are pursuing overlapping markets.

Unitree Robotics competes primarily with bipedal humanoids (H1, G1 — confusingly, Unitree also has a product called G1) at significantly lower price points. Unitree's G1 bipedal humanoid is priced around $16,000, a fraction of Galbot's wheeled G1. However, Unitree's manipulation capability and commercial deployment depth are less developed than Galbot's at the time of writing.

UBTECH Robotics (Walker series) is a more established Chinese humanoid company with longer operational history, though its commercial deployment scale in industrial settings is less clearly documented than Galbot's.

Agilex Robotics and Dobot compete in the mobile manipulation space with lighter-duty platforms at lower price points, targeting research and lighter industrial tasks rather than the heavy-payload factory floor.

AgiBot (formerly Zhiyuan Robotics) is a direct architectural peer — wheeled mobile manipulator, dual arms, AI-driven — and has also raised significant capital. The two companies are competing for the same automotive and logistics customers.

Competitive Positioning Table

Galbot's Differentiation Claims

Galbot's stated differentiators are: pure vision-based navigation (no floor markers or QR codes required) ²²⁰; a proprietary VLA model stack trained on synthetic data at scale ¹⁹²²; a 1-day deployment cycle into new retail environments ³; and the S1's autonomous battery swapping for continuous operation ²⁵. Each of these, if they perform as claimed in real commercial conditions, represents a meaningful operational advantage. The navigation claim in particular — centimetre-level positioning without infrastructure modification — is technically ambitious and partially supported by the AllDayNav paper ²¹, though that paper's test environments may not fully represent the diversity of real deployment sites.

The $800M–$930M in total funding ⁷¹¹¹⁵ gives Galbot a capital runway that most competitors cannot match, enabling simultaneous investment in hardware iteration, software model training at scale, and customer deployment support. This financial position is a genuine competitive advantage in a capital-intensive industry where the cost of training large embodied AI models is substantial.

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

Galbot is a Chinese company, founded in Beijing, funded substantially by Chinese state-backed capital, and deploying its products primarily within China. This context shapes its opportunities, its constraints, and the analytical lens through which Western buyers and investors should evaluate it.

State Capital and Strategic Alignment

The most recent disclosed funding round — 2.5 billion yuan (~$362M) led by the National AI Industry Investment Fund ⁷¹⁰ — is not ordinary venture capital. The National AI Industry Investment Fund is a state-directed vehicle; its participation signals that Galbot is considered strategically important to China's industrial policy agenda, specifically the push to automate manufacturing and reduce dependence on human labour in sectors like automotive and battery production. Other investors in this round include the Bank of China, CITIC Investment Holdings, and the China Integrated Circuit Industry Investment Fund ¹⁰ — all entities with direct or indirect state connections.

This funding structure has several implications. First, Galbot is unlikely to face capital starvation in the near term regardless of commercial performance, because its investors have non-commercial objectives (industrial policy, technology sovereignty) alongside financial returns. Second, the company's strategic direction will be influenced by state priorities, which currently emphasise domestic industrial deployment over international expansion. Third, Western governments and corporations evaluating Galbot as a supplier or partner will need to navigate the same due-diligence questions that apply to any Chinese technology company with state-backed investors: data governance, supply chain security, and potential export control exposure.

Export Controls and Technology Transfer

The G1 and S1 incorporate NVIDIA Jetson Thor compute in the premium variant ⁵. NVIDIA's export of advanced AI chips to China has been subject to evolving US restrictions, and the availability of Jetson Thor for Chinese commercial robotics applications may be affected by future regulatory changes. If NVIDIA hardware becomes restricted, Galbot would need to substitute domestic compute — likely from Huawei's Ascend line or similar — which could affect performance on the AI inference workloads that underpin the VLA models. The standard G1 uses Galbot's proprietary AstraBrain system ¹³, which reduces but does not eliminate this dependency.

The research collaboration network — Tsinghua University, Peking University, BAAI, CASIA, USTC, Shanghai AI Lab ^19202122 — is largely domestic, which limits one vector of technology transfer concern but also means the company's research base is subject to the same regulatory environment as its commercial operations.

International Market Access

Galbot's appearance at IREX Japan 2025 ²⁴ is the clearest signal of international market ambition. Japan is a logical first international target: it has a severe labour shortage in manufacturing and logistics, a sophisticated industrial customer base, and a cultural openness to automation that many Western markets lack. However, Japan also has stringent product safety certification requirements, and the G1's wheeled semi-humanoid form factor will require navigation through regulatory frameworks that do not yet have clear humanoid robot categories.

Entry into US or European markets faces additional friction. The Committee on Foreign Investment in the United States (CFIUS) scrutinises Chinese technology investments and deployments in sensitive sectors. European regulators are developing AI Act compliance frameworks that will apply to autonomous robots operating in commercial settings. Neither of these is an insurmountable barrier, but both add cost and timeline to international expansion.

The planned Hong Kong Stock Exchange IPO ⁷¹¹ is notable in this context. A HKEX listing gives Galbot access to international capital markets while remaining within a jurisdiction that, post-2020, operates under Chinese national security law. It does not provide the same international legitimacy signal as a NYSE or LSE listing, but it does create a public disclosure obligation that will improve the quality of independently verifiable information about the company's financials and operations.

Domestic Policy Tailwinds

China's "Made in China 2025" successor policies and the 14th Five-Year Plan both explicitly target robotics and intelligent manufacturing as priority sectors. The government's willingness to deploy state capital directly into Galbot — rather than simply providing subsidies or preferential procurement — suggests a high-confidence bet on the company's role in this agenda. CATL's participation as both investor and customer ¹⁶ is consistent with a coordinated industrial policy push to automate battery manufacturing as China seeks to maintain its dominance in EV supply chains.

This policy environment creates a protected domestic market in which Galbot can scale deployments, accumulate operational data, and iterate on its models before facing international competition. It is a structural advantage that Western competitors cannot replicate.

---

11The Hype, the Real and the Ugly

Galbot's public communications sit at the intersection of genuine technical achievement and promotional excess. Separating the two requires applying consistent evidentiary standards to each claim category.

What Is Demonstrably Real

The company exists, is well-funded, and has deployed robots in named commercial settings. The CATL partnership is confirmed by multiple independent news sources ⁶⁸¹² and by CATL's own public statements. The pharmacy deployments in Beijing are reported consistently across sources ³²⁶. The research publications are peer-reviewed and available on arXiv, with named co-authors from credible institutions ^19202122. The funding rounds are reported by Caixin Global, The Robot Report, and TechNode — outlets with editorial standards ⁷⁹¹¹¹³. The G1's physical specifications are consistent across multiple commerce and review sources ¹²³⁴.

The research results are the most independently verifiable element of the technical claims. The MobileH2R paper ¹⁹ demonstrates a +15% improvement in handover task success over baselines using synthetic-data-trained policies, tested on a real robot. The AllDayNav paper ²¹ demonstrates lifelong navigation in real environments using reinforcement learning. The LDA-1B paper ²² describes a 1-billion-parameter latent dynamics model trained on 8 million cross-embodiment samples. These are substantive contributions, not press release science.

What Is Claimed but Unverified

The "24/7 autonomous operation for over a year in warehouses and pharmacies" claim ³ is the most consequential unverified assertion. It is commercially significant — it is the claim that justifies the price point and the deployment model — but it rests entirely on vendor-sourced information. No independent operator has publicly confirmed it. No failure rate, downtime figure, or intervention frequency is disclosed.

The "95–97% grasp success rate" ³ is a vendor marketing figure. The research papers report relative improvements over baselines, not absolute success rates in commercial conditions. The two are not equivalent. A 97% success rate on a curated test set in a lab is not the same as 97% success across 5,000 SKU types in a live pharmacy.

The "5,000+ types of goods" handling claim ³ is similarly unverified. It may reflect the training data distribution rather than validated real-world performance across that full range.

The "1-day deployment into new stores" claim ³ lacks specification of what "deployment" means in this context. If it means the robot can navigate a new store layout after a single mapping pass, that is plausible given the AllDayNav framework ²¹. If it means full task capability across all SKUs in a new store within 24 hours, that is a much stronger claim requiring independent validation.

The Battery Life Conflict

The unresolved conflict between the "10+ hours" vendor claim and the "2 hours" aggregator figure ²³ is not a minor discrepancy. Battery life is a critical operational parameter for the 24/7 autonomy claim. If the G1 requires recharging every 2 hours, continuous operation requires either a charging infrastructure or multiple units in rotation — neither of which is mentioned in the deployment descriptions. The S1's stated 8-hour battery life ⁸ and autonomous battery swapping capability ²⁵ suggest Galbot is aware of this constraint and has addressed it in the heavier-duty variant, which may imply the G1's battery life is closer to the lower figure than the vendor claims.

The Ugly: What Is Not Disclosed

Several categories of information that would be material to any serious buyer or investor are entirely absent from public sources:

Failure modes and intervention rates. No source discloses how often the G1 requires human intervention during a shift, what failure modes are most common, or what the mean time between failures is. This is standard information for industrial automation equipment and its absence is notable.

Customer economics. No source discloses the total cost of ownership, including installation, maintenance, software licensing, and operator training. The unit price ($87K–$120K) is only one component of the economic case.

Regulatory certification. No source confirms CE marking, UL certification, or equivalent safety certification for the G1 or S1 in any jurisdiction. For a robot operating in proximity to humans in pharmacies and factories, this is a material gap.

Software update and support terms. The VLA models require ongoing training and updating as environments change. No source discloses the terms under which Galbot provides model updates, or what happens to deployed units if the company's commercial situation changes.

Actual order fulfilment. The "several thousand units ordered" figure ¹¹ is an order figure, not a shipment figure. The distinction matters enormously in an industry where announced orders have historically preceded delivery by years.

Claim tracker

Galbot G1/S1 operates fully autonomously (24/7, without teleoperation) in commercial deployments including CATL battery production lines, warehouses, and pharmacies for over a year.Unknown

Vendor and affiliated research sources consistently claim autonomous operation, and peer-reviewed papers (e.g., [19][21]) demonstrate closed-loop real-world task execution without teleoperation in controlled settings, but no independent third-party audit or user report in the dossier confirms uninterrupted 24/7 commercial autonomy at scale.

The Galbot G1 achieves a 95–97% grasp success rate and handles 5,000+ types of goods.Not supported

The 95–97% figure and 5,000+ goods count appear only in vendor/reseller listings [1][3]; the sole independent research validation [19] reports a relative +15% improvement over baselines in handover tasks but does not confirm any absolute grasp success rate figure.

The Galbot S1 heavy-duty variant supports 50 kg dual-arm payload, autonomous battery swapping, and 8-hour battery life for continuous industrial operation.Unknown

These specs are consistently reported across multiple news and video sources [8][12][25], including coverage by Interesting Engineering and Gasgoo, but all trace back to Galbot/CATL announcements with no independent hardware test or teardown confirming real-world payload or runtime performance.

Galbot's AI navigation system (AllDayNav) achieves near-100% navigation success in real-world lifelong deployment using reinforcement learning without human intervention.Supported

The AllDayNav framework is documented in a peer-reviewed arxiv paper [21] co-authored with multiple independent academic institutions (Tsinghua, BAAI, etc.), demonstrating real-world RL-based navigation without teleoperation; however, the specific success-rate figure and generalization beyond tested environments remain unverified by external replication.

The Galbot G1 has a battery life of 10+ hours, enabling extended autonomous operation.Not supported

The 10+ hour claim appears in a commerce/reseller listing [1][4] and aligns with vendor marketing, but a separate aggregator source [2] lists only 2 hours runtime; neither figure comes from an independent test, and the dossier explicitly flags this as an unresolved conflict — making the headline 10+ hour claim unsupported.

Galbot's GraspVLA and related VLA models are trained entirely on synthetic data (sim-to-real transfer) without real-world demonstrations, yet achieve real-world task performance.Supported

Peer-reviewed research [19][22] co-authored with independent institutions (Tsinghua, BAAI, USTC, Shanghai AI Lab, NVIDIA, etc.) explicitly documents synthetic-data-only training with successful real-world deployment on the G1 hardware, providing independent academic corroboration of this methodology — though broader generalization limits are not fully characterized.

---

12Future Scenarios

The following scenarios are editorial inferences from the available evidence. They are not predictions. Each is assigned a rough probability range based on the strength of the supporting evidence.

Scenario A: Scaled Domestic Industrial Deployment (Most Likely, ~55% probability)

Galbot successfully delivers on its CATL, Bosch, and Toyota orders over the next 18–36 months, accumulates operational data from thousands of deployed units, and uses that data to improve its VLA models iteratively. The Hong Kong IPO proceeds, providing additional capital and a public disclosure regime. The company becomes the dominant supplier of semi-humanoid mobile manipulators to Chinese automotive and battery manufacturers, with a defensible position built on proprietary operational data and deep customer integration.

The risks to this scenario are execution-related: manufacturing scale-up, quality control at volume, and the challenge of maintaining software performance across a diverse installed base. The CATL investor-customer relationship mitigates some commercial risk but creates dependency.

Scenario B: International Expansion Gains Traction (Plausible, ~25% probability)

The IREX Japan 2025 appearance ²⁴ converts into commercial deployments with Japanese automotive or logistics customers. The HKEX IPO attracts international institutional investors who provide both capital and market access. Galbot navigates export control and safety certification requirements in Japan and potentially Southeast Asia, establishing a beachhead outside China before attempting European or North American markets.

This scenario requires Galbot to solve regulatory and certification challenges that are not currently addressed in public disclosures. It also requires the G1's performance to hold up in environments with different physical characteristics (older factory floors, different shelf standards) and different operational cultures.

Scenario C: Technology Commoditisation Erodes Moat (Moderate Risk, ~30% probability over 3 years)

Foundation model capabilities for robot manipulation improve rapidly across the industry, reducing the differentiation value of Galbot's proprietary VLA stack. Open-source or widely licensed manipulation models (from Physical Intelligence, Google DeepMind, or Chinese equivalents) reach performance parity with GraspVLA and GroceryVLA. Galbot's competitive position then depends more heavily on hardware cost, manufacturing scale, and customer relationships — areas where it faces intense domestic competition from AgiBot, UBTECH, and others.

This scenario does not necessarily mean Galbot fails; it means the company's margin structure and valuation multiple come under pressure as software differentiation erodes.

Scenario D: Geopolitical Disruption (Low-to-Moderate Risk, ~20% probability)

Escalating US-China technology restrictions affect NVIDIA chip availability for the G1 Premium variant ⁵, forcing a hardware redesign. Alternatively, Western customers (Bosch, Toyota's non-Chinese operations) face political or regulatory pressure to reduce procurement from Chinese robotics suppliers with state-backed investors. The HKEX IPO is delayed or withdrawn due to market conditions or regulatory complications.

The state-backed funding structure that protects Galbot from capital starvation also makes it a more prominent target for technology restriction measures. The domestic compute fallback (AstraBrain) provides partial insulation but at potential performance cost.

Scenario E: Autonomy Claims Prove Overstated (Low Probability but High Impact, ~15% probability)

Independent testing or a high-profile deployment failure reveals that the G1's real-world autonomy performance is significantly below vendor claims — that intervention rates are high, that the "24/7" figure refers to availability rather than unassisted operation, or that the 95–97% grasp success rate does not hold across the full claimed SKU range. This would not necessarily be fatal — many industrial automation products require more human oversight than initially marketed — but it would damage the premium pricing justification and require a recalibration of the commercial model.

The research publications ¹⁹²¹²² provide some protection against this scenario: they demonstrate real capability in controlled conditions. The gap between controlled research conditions and messy commercial reality is the key uncertainty.

Scenarios are not mutually exclusive. A and C could occur simultaneously; B and D could interact.

---

13What to Watch: A Live Monitoring Checklist

The following indicators, if they become publicly available, would materially update the analysis in this report. Analysts and procurement teams should monitor these signals on a rolling basis.

Commercial Validation Signals

• Named customer confirmation of G1 deployments beyond CATL and pharmacies. Bosch and Toyota are named as order-placers ¹¹ but have not publicly confirmed deployments. A press release or case study from either would upgrade the commercial evidence tier significantly.
• Shipment figures versus order figures. The "several thousand units ordered" ¹¹ needs to be matched against actual delivery and installation data. Watch for IPO prospectus disclosures, which will require revenue and unit shipment figures.
• Independent operator testimony. Any pharmacy chain, warehouse operator, or automotive supplier that publicly discusses G1 performance — including failure rates, intervention frequency, and total cost of ownership — would be the most valuable single data point currently missing from the public record.

Technical Performance Signals

• Battery life independent test. The unresolved conflict between 2-hour and 10+-hour figures ²³ needs resolution. Any independent teardown, user report, or third-party review that tests runtime under realistic load conditions should be tracked.
• Regulatory certification announcements. CE marking, UL listing, or equivalent safety certification in any jurisdiction would confirm the G1 is cleared for the human-proximate environments it is deployed in.
• New research publications. The affiliated research group is prolific ^19202122. New papers on manipulation success rates in uncontrolled environments, long-horizon task completion, or failure mode analysis would update the technical evidence base.
• Open-source model or dataset releases. The EI-30k dataset ²¹ and LDA-1B model ²² are described in papers but not confirmed as publicly released. Any public release would allow independent evaluation of the underlying capabilities.

Financial and Corporate Signals

• HKEX IPO filing. The prospectus will contain audited financials, revenue breakdown by customer and geography, unit economics, and risk disclosures. This is the single most important forthcoming document for independent analysis.
• Additional funding rounds. The pace of fundraising (multiple rounds in 2025–2026 ^79111315) suggests ongoing capital needs. The terms and investors of future rounds will signal whether commercial revenue is scaling or whether the company remains primarily investment-funded.
• CATL investor-customer relationship evolution. Watch for any change in CATL's stake, any exclusive supply agreement, or any indication that the CATL deployment is being scaled to additional facilities beyond Luoyang.

Competitive and Market Signals

• AgiBot deployment announcements. As the closest domestic architectural peer, AgiBot's commercial progress is the most direct benchmark for Galbot's competitive position.
• Agility Robotics Amazon deployment scale. If Amazon publicly discloses Digit deployment numbers or performance data, it will provide a Western reference point against which Galbot's claimed metrics can be compared.
• Chinese government procurement mandates. Any policy directive requiring state-owned enterprises to procure domestically produced humanoid robots would be a significant demand catalyst for Galbot given its investor base.
• Export control developments. Monitor US Bureau of Industry and Security (BIS) rule changes affecting NVIDIA Jetson Thor availability in China, and any CFIUS guidance on Chinese robotics company operations in Western markets.

Red Flags to Watch

• Withdrawal or delay of the HKEX IPO without explanation.
• Any public report of a significant G1 deployment failure, injury incident, or customer contract termination.
• Departure of founder Wang He or key research leadership from affiliated institutions.
• Reduction in research publication output from the affiliated academic network, which would suggest either IP consolidation (positive) or research capability decline (negative).
• Any revision downward of the "several thousand units ordered" figure, or silence on order fulfilment timelines in IPO disclosures.

---

14Sources and Methodology

Methodology

This report was produced from a structured research dossier compiled on 25 June 2026, comprising 33 numbered sources across six categories: official company materials (0), commerce/product listings (5), research publications (4), news articles (13), video content (6), and community sources (5). The overall dossier confidence score assigned by the compilation process was 0.72.

Evidence classification. All claims in this report are classified into one of four categories: VERIFIED FACT (confirmed by regulatory filings, official product documentation, named-customer confirmation, peer-reviewed research, or multiple independent sources); COMPANY CLAIM (stated by Galbot or its affiliates, not independently verified); EDITORIAL INFERENCE (reasoned conclusions drawn from the available evidence); or UNKNOWN (not publicly disclosed). These classifications are applied inline throughout the report.

Source quality weighting. Peer-reviewed arXiv preprints with named institutional co-authors are treated as higher-quality evidence than commerce aggregator listings or YouTube video descriptions, but lower than published journal articles or regulatory filings. News articles from outlets with editorial standards (Caixin Global, The Robot Report, TechNode) are treated as reliable for factual claims about funding rounds and named partnerships, but not for technical performance claims that originate with the company.

What this report does not do. This report does not independently test any Galbot product. It does not conduct primary interviews with company personnel, customers, or competitors. It does not access non-public financial information. All conclusions are drawn from publicly available sources as listed below.

Community sources excluded. Sources ²⁸ through ³³ in the dossier — comprising an unrelated YouTube review, and Reddit threads on unrelated topics (cars, philosophy, cryptocurrency, van conversions, and network security) — contain no material information about Galbot and are excluded from citation throughout this report. Their presence in the dossier appears to reflect noise in the automated research compilation process.

Sources

¹ Galbot G1 Wheeled Humanoid Robot | Robots International — https://www.robotsinternational.com/Galbot-G1.htm

² Galbot - Humanoid.guide — https://humanoid.guide/product/galbot

³ Galbot G1 Review: Price ($87K) & Specs ²⁰²⁶ | Robozaps — https://blog.robozaps.com/b/galbot-g1-review

⁴ Galbot G1 Humanoid Robot — Buy or Lease | SVRC — https://www.roboticscenter.ai/store/product/galbot-g1

⁵ Galbot G1 Premium: The Swiftest Humanoid Robot Worker, Powered by NVIDIA Jetson Thor — https://www.youtube.com/watch?v=SgmhntLZaEE

⁶ CATL teams with Galbot to scale humanoid robots in battery factories — https://cnevpost.com/2026/06/24/catl-galbot-scale-humanoid-robots-battery-factories/

⁷ Galbot Raises $362 Million in Fresh Funding, Eyes Hong Kong IPO - Caixin Global — https://www.caixinglobal.com/2026-03-03/galbot-raises-362-million-in-fresh-funding-eyes-hong-kong-ipo-102418742.html

⁸ CATL's battery-making factory gets Galbot's humanoid as new worker — https://interestingengineering.com/ai-robotics/catl-battery-powered-heavy-load-humanoid-robot-factory

⁹ Galbot brings in $300M to scale mobile manipulator deployments - The Robot Report — https://www.therobotreport.com/galbot-brings-in-300m-to-scale-mobile-manipulator-deployments/

¹⁰ Chinese robotics startup Galbot secures major state backing to scale AI models - CnTechPost — https://cntechpost.com/2026/03/02/galbot-secures-major-state-backing/

¹¹ Humanoid robot maker Galbot raises RMB 2.5 billion · TechNode — https://technode.com/2026/03/02/humanoid-robot-maker-galbot-raises-rmb-2-5-billion/

¹² Gasgoo Daily: CATL, Galbot deploy first heavy-duty humanoid robot powered by CATL batteries | Gasgoo — https://autonews.gasgoo.com/articles/video/gasgoo-daily-catl-galbot-deploy-first-heavy-duty-humanoid-robot-powered-by-catl-batteries-2069784496696958976

¹³ Galbot picks up $153M to commercialize G1 semi-humanoid - The Robot Report — https://www.therobotreport.com/galbot-picks-up-153m-commercialize-g1-semi-humanoid/

¹⁴ Robotics Company Raises $150M for Embodied AI Robot Deployment — https://aibusiness.com/automation/robotics-company-raises-150m-for-embodied-ai-robot-deployment

¹⁵ Seeds | GALBOT Completes New Funding Round, Raising Nearly 5 Billion Yuan in Three Months | Gasgoo — https://autonews.gasgoo.com/articles/news/seeds-galbot-completes-new-funding-round-raising-nearly-5-billion-yuan-in-three-months-2028713877083000832

¹⁶ CATL backs humanoid robotics startup Galbot in $153 million funding · TechNode — https://technode.com/2025/06/25/galbot

¹⁷ 6 Billion Yuan in Funding! Galbot and Daimon Robotics Secure Major Investments — https://cnmra.com/6-billion-yuan-in-funding-galbot-and-daimon-robotics-secure-major-investments

¹⁸ Galbot picks up $153M to commercialize G1 semi-humanoid - The Robot Report — https://www.therobotreport.com/galbot-picks-up-153m-commercialize-g1-semi-humanoid

¹⁹ MobileH