Galbot Market Research

1. Introduction

Galbot's differentiator within this group is "retail-scenario deployment": it is the only one among the ten companies here that has turned unmanned retail and smart pharmacies into a scaled self-operated scenario — its "Galaxy Pod" stores expand at a density of "two-day deployment, one unit managing 5,000+ SKUs," and it obtained the world's first drug business license issued for a humanoid robot. Technically, it chose the combination of a "wheeled bi-arm body + full-stack in-house embodied foundation model"; it does not develop its own body (procuring from Unitree and others, contributing over 18 million RMB in purchases to Unitree in 2025), concentrating resources on algorithms and data.

1.1. Company Overview

1.2. Development History and Funding Cadence

1.3. Business Model and Shipment Footing

Galbot's revenue structure runs on three lines — "solution delivery + self-operated retail scenarios + industrial bulk orders": it delivers S1 solutions to industrial customers (CATL, Baida Precision, etc.); it operates its own Galaxy Pod retail / pharmacy network to validate and feed back into the model; and the home robot ET1's pre-sale opens the consumer line. Shipments are not officially disclosed (third-party estimates of several hundred units), differing from Zhiyuan / Unitree's multi-thousand-unit shipments — Galbot's per-unit value is high (G1 ~700,000 RMB, S1 ~1,050,000 RMB), and "few units, high unit price, heavy solutions" is its commercial profile.

2. Glossary

3. Function Description and Product Lines

3.1. Product Line Overview

3.2. Core Models and Algorithms

1. AstraBrain full-brain system: brain (WAM 0.5, a general-purpose brain foundation unifying VLA and the world model), cerebellum (WBC 0.5, with Scaling Law officially validated), neural control, and the AstraBrain-Dex dexterous-manipulation world model.
2. GraspVLA: grasping foundation model, a representative achievement of the synthetic-data pre-training route.
3. WAM-TTT: released 2026-07, a post-training paradigm requiring no human action labels; the official claim is that it "doesn't lose its craft when the kitchen changes" — aimed squarely at the real-robot data generalization problem.
4. NavFoM + DexNDM: two capability foundations for navigation and dexterous hands.
5. Proof of motor capability: 2026-03 world's first humanoid robot to play tennis autonomously (ball recognition, trajectory prediction, whole-body balance, consecutive shots); 2026-08 gold medals in the home / catering / retail, the three hardest events at the sports meet.

4. Platform Architecture

4.1. Galaxy Brain: The AstraBrain Whole-Brain Architecture

AstraBrain is explicitly described on the official site as "the world's first whole-body end-to-end embodied foundation model integrating brain, cerebellum, and neural control in a single model." The division of labor has three layers: the brain model (WAM) handles semantic reasoning and world understanding; the cerebellum model (WBC) handles whole-body motion control; the neural-control layer handles low-latency reflexes; and AstraBrain-Dex targets dexterous manipulation. This layering parallels the System 0/1/2 three-tier system of LimX Dynamics (see 07-robotera.md), but every layer of Galbot's stack has a clear model product name and release cadence.

4.2. Synthetic-Data-Dominated Training Pipeline

Galbot is the vendor in this group that pursues the synthetic-data route most thoroughly: billion-scale simulated synthetic-data pre-training (GraspVLA) + real-robot fine-tuning, forming a three-way contrast of data routes with Zhiyuan (primarily real-robot data-collection factories) and Figure (first-person human video). WAM-TTT in 2026-07 further reduces the dependence on human action labels — the combination of synthetic data plus self-supervised post-training aims to resolve the contradiction between data cost and scenario generalization at the same time.

4.3. AstraBrain Three-Tier Integrated Structure Diagram

Fig. 6-1 | Galaxy Brain: The AstraBrain Three-Tier Integrated Architecture (schematic)

Schematic: drawn from this document's analysis of AstraBrain's official architecture; model names and positioning are taken from the Galbot official site and 36Kr reports.

5. Harness Design

5.1. L1 Context Engineering Layer

WAM 0.5 unifies visual-linguistic understanding with world-model prediction; task context (scene, objects, instructions) is carried internally by the world model, while NavFoM supplies spatial context. Synthetic-data pre-training exposes the model to far more environmental variation than real deployments, which is equivalent to pre-loading the "context prior" during training. No explicit retrieval / compression components are disclosed.

5.2. L2 Tools and Execution Layer

Dual-arm execution + DexNDM dexterous-hand neural dynamics + WBC whole-body control form the execution stack. Autonomous tennis and the three gold medals at the sports meet are the strongest public evidence of dynamic performance at the execution layer; the S1's 50kg dual-arm load covers industrial heavy-load work. This layer has ample evidence and receives the highest rating.

5.3. L3 Orchestration and Control Layer

The Galaxy Pod stores' "single unit managing 5,000+ SKUs" requires continuous multi-task orchestration: restocking, replenishing, pharmacy-sales interaction, and anomaly reporting are autonomously queued and executed unattended. The 7×24 normal operation (CATL) likewise depends on task-level long-horizon orchestration and interruption recovery. The engineering form of the orchestration framework is not public, but its operating results provide indirect proof of L3's effectiveness.

5.4. L4 Memory and State Layer

WAM-TTT (Test-Time Training) is essentially a runtime experience-injection mechanism: the robot updates its capabilities through test-time training in a new environment, equivalent to pushing "experience learning" down to the deployment site. This is the most methodologically innovative L4 material in this group. In-store inventory-state and order-state management belong to the conventional state layer.

5.5. Evaluation and Observation Layer

Three sets of evidence: gold medals in the three hardest events at the sports meet (a comparable competitive benchmark); more than 3 months of continuous 7×24 operation at CATL (operations-level stability observation); and scaled operational data from 170+ / 100+ stores (two figures). There is no public standardized evaluation benchmark, but the density of operations-based evaluation evidence leads this group.

5.6. L6 Governance and Safety Layer

Two hard pieces of evidence: the world's first drug business license issued for a humanoid robot (regulatory access for the pharmacy-sales scenario); and the global strategic cooperation with CATL's Ningjia Service (1,300+ service stations handling deployment and maintenance), outsourcing operation, maintenance, and safety response to a national-level service network. Leading approval of the key laboratory for embodied-intelligence foundation models provides research-side endorsement on data and model governance. Privacy governance details for the home robot ET1 are not disclosed.

5.7. Six-Layer Strength Summary

6. Real-World Cases

6.1. CATL Mass-Production Line 7×24 Normal Operation

Reported 2026-06 (by industry media such as Embodied Research Society): the Galbot S1 achieved more than 3 months of continuous 7×24 operation at CATL's battery plant, handling long-horizon autonomous tasks such as module and battery-pack material transfer, described as the world's first normalized autonomous operation of a humanoid robot in new-energy manufacturing. Galbot has entered real production lines at CATL, Bosch, Siemens, Toyota, Hyundai, BAIC, SAIC, Zeekr, Great Wall, and others. It should be noted: this case relies primarily on a single industry-analysis article and 36Kr re-publication; official press-release-level sources remain to be reinforced (see the Information-Gap Statement). The order of over 1,000 units signed with Baida Precision in 2025-12 is a representative contract for industrial batch scaling.

6.2. Galaxy Pod Unmanned Convenience Stores and Smart Pharmacies

As of 2026-03, Galaxy Pod stores had been rolled out in more than 20 cities: the 36Kr figure is more than 170 and the SPDB International figure is more than 100 — both figures are shown side by side, without choosing one. Each store has a two-day deployment cycle; a single robot manages 5,000+ SKUs; and the pharmacy format holds the world's first drug business license issued for a humanoid robot. The deeper value of the Galaxy Pod is data and scenario sovereignty: because the stores are self-operated, Galbot is simultaneously supplier and operator, and task distributions, failure modes, and user-interaction data are all internalized. The 2026-02 Spring Festival Gala holiday micro-film (in which the G1 shared the stage with Shen Teng and Ma Li, completing actions such as shelling walnuts, picking up glass, folding clothes, and skewering sausages) delivered national-scale exposure.

6.3. Going Global and the Operations Network

It formed the joint venture "Boyin Hechuang" with Bosch (2025-06), plans to debut at Bosch's BCW conference in Germany in 2026-06, and enters the European market through the joint-venture model; the global strategic cooperation with CATL's Ningjia Service uses 1,300+ service stations to take on deployment and maintenance, addressing the after-sales weakness of scaled humanoid-robot deployment — this is the only systematic third-party operations-network arrangement among the vendors in this group.

7. Summary

7.1. Strengths and Limitations

Strengths: the synthetic-data route keeps data costs significantly lower than real-robot data collection; the AstraBrain whole-brain system is complete in layers and iterates quickly (three levels up in half a year across WAM / WBC / WAM-TTT); the Galaxy Pod self-operated scenarios provide a unique operational-data loop; and it benefits from the dual backing of national capital (National Big Fund Phase III's first investment) and CATL industrial capital. The L6 evidence (drug license + operations network) is the strongest in this group.

Limitations: the body is not self-developed, so hardware scaling is constrained by the supply chain (Unitree and others); shipment volume is small (an estimated few hundred units), not on the same scale as vendors shipping thousands; the three valuation figures diverge widely (20 / 21 / 26 billion RMB); there is no official first-hand press release for the CATL case; and the ET1 home line is still in its pre-sale validation phase.

7.2. Applicable Boundaries

It fits service scenarios centered on grasping, transporting, and interacting that do not require bipedal obstacle traversal (retail, pharmacy, supermarket — where the wheeled bi-arm is the most efficient solution), as well as industrial scenarios involving heavy-load material transfer such as batteries and automotive parts. Bipedal all-terrain and complex obstacle-traversal scenarios are beyond its body's capability; budget-sensitive research and education customers should choose Unitree's lower-priced tier.

7.3. Selection Recommendations

1. Retail / pharmacy customers should prioritize examining the Galaxy Pod's existing operational data (SKU management density, replenishment cadence, anomaly rates) rather than demo videos.
2. Industrial customers should demand official confirmation and a site visit for the CATL case, and use the delivery cadence of the Baida Precision bulk order as a reference.
3. When purchasing the G1/S1, agree in parallel on a Ningjia Service or equivalent maintenance SLA — the operations network is part of the value of this solution.
4. Follow WAM-TTT's rollout: if "scenario switching without labels" matures, it will significantly reduce the implementation cost of secondary deployments.

Information-Gap Statement

1. Conflict among the three valuation figures: ~20 billion (overseas popular-science article) / 21 billion (36Kr) / 26 billion RMB (SPDB International); this document presents them side by side without choosing one.
2. Conflict between the two Galaxy Pod store figures: more than 170 (36Kr / industry-analysis article) vs more than 100 (SPDB International); presented side by side.
3. Shipment volume is not officially disclosed; third-party estimates are on the order of several hundred units, for scale reference only.
4. The CATL 7×24 normalized-operation case relies mainly on an industry-analysis article and 36Kr re-publication; an official press release is pending reinforcement.
5. ET1's pre-sale price and technical specifications are not fully disclosed, marked [To be filled].
6. Musk's comment on the tennis match is an unsourced re-publication (grade D); this document does not adopt it.

8. References

1. Galbot official website (AstraBrain / WAM / WBC / Dex architecture) — Galbot, 2026. https://www.galbot.com
2. Galbot news (Galaxy Xingzi reservations / three gold medals at the sports meet / laboratory approval) — Galbot, 2026. https://www.galbot.com/news
3. The world's first native-agent robot "Galaxy Xingzi" hotly accepting reservations — Galbot Robotics official WeChat account, 2026-09-03. https://mp.weixin.qq.com/s/RnmqpkT5Y-DNg4QE9ga1bg
4. 100% gold-medal rate! Galbot sweeps the three hardest event gold medals (home, catering, retail) — Galbot Robotics official WeChat account, 2026-08-26. https://mp.weixin.qq.com/s/QQs2y7jEzPDkWHAt_GFUEw
5. China's only key laboratory for embodied-intelligence foundation models, led by Galbot, is approved — Galbot Robotics official WeChat account, 2026-08-10. https://mp.weixin.qq.com/s/6IdDqb0yxYQU7eQKAsolDw
6. Upending the post-training paradigm! WAM-TTT lets robots not lose their craft when the kitchen changes — Tencent Technology, 2026-07-16. https://mp.weixin.qq.com/s/LMtt3FGB6zyB7Boex4S6kg
7. Exclusive | The key to CATL's full embodied-intelligence chain has been handed to Galbot — Embodied Research Society, 2026-06-25. https://mp.weixin.qq.com/s/VsQ2Uir_Vn1EF40Bc0xGaQ
8. Galbot: 34 months since founding, 6 rounds of funding totaling 7 billion — 36Kr, 2026. https://36kr.com/p/3978672823417862
9. Galbot Robotics entry — Baidu Baike, 2026. https://baike.baidu.com/item/%E9%93%B6%E6%B2%B3%E9%80%9A%E7%94%A8%E6%9C%BA%E5%99%A8%E4%BA%BA/68609336
10. SPDB International research report (AI from digital to physical: the Galbot / accelerated evolution chapter) — SPDB International, 2026. https://ima.qq.com/wiki/?shareId=5fe3cdeeeef912a0fc6bac2c6c801c9d3f4da509fb4ff731aed83df7c6bd810c