AgiBot Market Research

1. Introduction

AgiBot's intra-group differentiation thread is "Data Collection Factory + WholeBodyVLA full stack": it is the most vertically integrated vendor among the ten in this group, covering at the same time the embodiment, embodied foundation model, world model, simulation platform, operating system, and open-source datasets, and it is also the first vendor in China to build its own dedicated data collection factory. On the Omdia metric, it shipped 5,168 units in 2025 with roughly 35% global market share, forming the shipment leader together with Unitree (see 01-unitree.md for details).

1.1. Company Overview

1.2. Development History and Key Milestones

1.3. Business Model and Capital Path

AgiBot's revenue structure runs three tracks in parallel: whole-machine sales + scenario solutions + data and development-tool ecosystem. Whole machines target industrial training, commercial performance, and research; Genie Studio / Lingqu OS / AgiBot World form an open-source ecosystem for developers that lowers the migration cost for industry customers; and on the data side the data collection factory carries both self-use training and potential external data services. On the capital path, the A-share market (controlling Shangwei New Materials) and a Hong Kong IPO (reported) advance on dual tracks, in contrast to Unitree's direct IPO and UBTech's Hong Kong-first approach. No official price list has been obtained for the full series of models; marked [To be filled].

2. Glossary

3. Feature Description and Product Lines

3.1. Product Line Overview

3.2. Core Models and Software Stack

1. GO series: GO-1 (ViLLA framework, officially claimed success rate 46%→78%) → GO-2 (2026-04, new-generation embodied foundation model).
2. World model line: Genie Envisioner (unified world model platform) → GE-Sim 2.0 (closed-loop training physics engine) → Genie Sim 3.0 (LLM-driven open-source simulation) → GE-Act 2.0 (2026-09, first validation of the native world-action model pretraining and Scaling route).
3. Development platform: Genie Studio (2025-04) → Genie Studio Agent (2026-04, no-code embodied job platform).
4. System layer: Lingqu OS + AimRT communication framework, open-sourced across the full stack.
5. Dataset: AgiBot World (million-robot real-robot dataset) → AGIBOT WORLD 2026; VideoDataset; the first CR certification for a humanoid-robot dataset.
6. Motion intelligence: AGILE2.0 perception-control integrated model (2026-09-11, officially said to usher robots into the "eyes-open movement" era).

4. Platform Architecture

4.1. Full-Stack Technology Landscape

AgiBot is the only vendor in this group with its own products — partially open-sourced — across all six levels of "embodiment — model — simulation — data — operating system — development tools." GenieReasoner (2026-01) integrates high-level semantic reasoning with low-level fine-grained control, improving inference accuracy by 25% per the official metric; WholeBodyVLA extends VLA from a fixed base to bipedal whole-body control. Against the Harness six-layer model, this landscape has the fullest coverage: GO series and WholeBodyVLA are the L1/L2 model stack, Genie Studio and Lingqu OS are the L2/L3 layers, AgiBot World + data collection factory + GE-Sim is the L4/L5 data-evaluation loop, and CR certification is the L6 compliance material.

4.2. Data Collection Factory and Data Feedback Loop

AgiBot built China's first dedicated professional data collection factory (3,000+ square meters, collecting 30,000–50,000 data items daily), and in 2026-05 activated the "1+N distributed data collection training ground" model: a central base + regional training grounds (the first located at the Chengdu Pidu Southwest Base). Data enters GO-series model training after collection, cleaning, and labeling, and then flows back through GE-Sim simulation evaluation and real-robot verification, forming an industrial-grade "real-robot data — model — real-robot deployment" closed loop. AgiBot World open-sources part of this loop's outcomes and is currently the world's largest open-source real-robot dataset (more than 10 times the scale of DROID).

4.3. Data Collection Feedback Loop Diagram

Figure 2-1 | The data-training-deployment closed loop driven by the AgiBot data collection factory

Schematic: drawn based on this document's analysis; key figures come from the Zheshang Securities report, CNR (China National Radio), and AgiBot official disclosure, with information cut off as of 2026-09-12.

5. Harness Design

5.1. L1 Context Engineering Layer

GO series + GenieReasoner form a perception-language-reasoning context assembly stack: high-level semantic reasoning is responsible for task-level understanding, and Genie Studio Agent's no-code job configuration determines how task context is injected. The AgiBot World dataset itself can also be viewed as "pretraining corpus for the embodied domain," available to external model vendors for post-training at L1/L2.

5.2. L2 Tools and Execution Layer

The embodiment families (Yuanzheng / Lingxi / Jingling / D1 / C5) + self-developed dexterous hands and motors form the execution hardware layer; Lingqu OS and AimRT provide runtime communication and tool registration mechanisms; Genie Sim 3.0 allows execution policies to be validated in simulation first. The open-source strategy (Lingxi X1 full stack, GO-1, AimRT) significantly lowers the integration cost for external developers at L2.

5.3. L3 Orchestration and Control Layer

WholeBodyVLA extends orchestration from chassis motion to bipedal whole-body control; GenieReasoner's integrated big-brain/small-brain design (high-level planning + low-level fine-grained control) naturally corresponds to L3's "plan-execute" layering. Public details on multi-robot coordination and task-level workflow orchestration are limited, so the rating is clear mechanism, engineering details not visible.

5.4. L4 Memory and State Layer

AgiBot's L4 evidence is on the data side rather than the runtime side: the AgiBot World million-robot dataset and VideoDataset form a cross-task "embodied experience library," and GE-Sim 2.0's trainable/evaluable engine supports experience-replay-style training. Runtime memory (session state, long-term memory) lacks public disclosure; marked [To be filled].

5.5. Evaluation and Observation Layer

This is AgiBot's notable strength: GE-Sim 2.0 is positioned as an "interactive, trainable, evaluable" closed-loop training engine; ACoT-VLA is open-sourced as the official baseline of the AGIBOT WORLD CHALLENGE, meaning AgiBot is building a public evaluation benchmark for the embodied domain; and being first in both the gold-medal and medal tables at the robot games (2026-08, official) provides comparable results competed on the same field as rivals. The ramp-up to 15,000 mass-produced units is itself an outcome of manufacturing-reliability observation.

5.6. L6 Governance and Safety Layer

The industry's first CR certification for a humanoid-robot dataset (No. 001, 2025-09) is concrete evidence at the data-compliance level. Runtime safety governance (job safety fencing, permission tiers, auditing) lacks public detail; the Yuanzheng A3's 10-second battery swap and Guinness-level continuous-walking reliability (106.286 km, hot-swap battery replacement without ever powering off) are availability, not governance, evidence.

5.7. Six-Layer Strength/Weakness Summary

6. Case Studies

6.1. Southwest Base and Distributed Data Collection Training Ground

On 2026-05-22, AgiBot's Southwest Base (Chengdu Pidu) and the first "1+N" distributed data collection training ground were activated, with the first 200 units rolling off the line; the Yuanzheng A3 is the core intelligent-manufacturing model, with an annual output of several thousand units at full capacity. This model expands the data collection factory from a single Shanghai site into a distributed network, directly supporting the ramp-up pace of "the 15,000th mass-produced unit rolling off the line" (2026-07, officially confirmed).

6.2. In-Factory Operation at Fulin Precision

The Yuanzheng A2-W has been deployed at the Fuling Jinggong factory, performing parts/pallet loading and unloading and power battery plug/unplug testing — a representative named-customer case in AgiBot's industrial scenarios, forming a comparable parallel to the automotive-factory training programs of Galaxy General (CATL) and UBTech (BYD, etc.).

6.3. Munich Launch Event and Minsheng Group Collaboration

On 2026-02-24 AgiBot held a launch event in Munich, Germany, entering the German market, and signed a strategic cooperation with Minsheng Group, with Minsheng's European factory serving as a training and data collection base. This is one of the earliest moves among Chinese vendors in this group to enter Europe's industrial data collection ecosystem under their own brand.

6.4. Extreme Reliability Validation and Ecosystem Layout

The Yuanzheng A2 set a Guinness World Record on 2025-11-20: continuous walking of 106.286 km with hot-swap battery replacement throughout and never powering off — the hardest public evidence of "runtime continuity." On the ecosystem side, AgiBot has established joint ventures or strategic cooperation with Lens Technology, BOPU Precision, Junpu Intelligent, and Lingyi iTech, among others; the Wuxi Experience Center, Zhuhai Embodied Intelligent Application Innovation Center, and Ningbo Innovation Center have been established; the first overseas venue landed in Malaysia (2026-01); and in 2026-08 it joined with Chimelong to build the world's first embodied-intelligent theme park.

7. Summary

7.1. Strengths and Limitations

Strengths: the highest degree of vertical integration in this group, with all six layers — model (GO series), world model (GE), simulation (Genie Sim), data (AgiBot World), system (Lingqu OS), and embodiment — complete and largely open-sourced; globally top-two shipments with the fastest ramp-up (5,000 → 15,000 units in about 7 months); and the data collection factory and distributed training grounds form a replicable data infrastructure.

Limitations: the first-place shipment position faces competition from Unitree (5,168 units vs Unitree's 5,500+ under the two metrics); the depth of industrial named-customer cases does not match UBTech (multiple customers, multiple scenarios, two years of training) or Galaxy General (CATL 7×24 ongoing operation); the listing path is still uncertain (reported); and full-series pricing is opaque.

7.2. Applicability Boundaries

Suitable for industrial customers and developers that need full-stack control + an open-source ecosystem: for teams that want to do secondary development based on open-source real-robot data and open-source simulation, AgiBot is the most complete supply option in this group. If you only need to buy embodiments for research and education, Unitree's price band is more favorable; if you want pure model licensing without embodiments, you should consider PI (see 08-physical-intelligence.md for details).

7.3. Selection Recommendations

1. For industrial-training-type requirements, prioritize evaluating the combination of Yuanzheng A3 (battery-swap continuity) and Lingxi X2 (lightweight scenarios).
2. Developer teams are advised to start from the Genie Studio + AgiBot World + GO-1 open-source stack, verifying algorithms before purchasing real machines.
3. Before purchasing, require the vendor to clarify the delivery metric: shipment volume (Omdia 5,168 units / IDC about 5,000 units) is an industry metric, and project delivery capability must be due-diligenced case by case by customer reference.
4. Watch the subsequent landing of GE-Act 2.0 and AGILE2.0: if the world-action model route is validated, it will change the dependency on simulation data.

Information Gap Statement

1. Full-series pricing lacks an official price list; marked [To be filled].
2. Two-metric conflict: 2025 shipments of 5,168 units (Omdia / Zheshang Securities) and about 5,000 units (IDC) are presented in parallel, without choosing one.
3. Hong Kong IPO (target valuation USD 5.1–6.4 billion, sponsors) is per media reports and not officially confirmed; this document notes it per the reported metric.
4. Runtime memory and security governance (L4 / L6) lack public detail.
5. The Yuanzheng A3 full parameter table has not been item-by-item verified; this document lists only the core parameters confirmed by the official site and reports.

8. References

1. AgiBot holds a launch event in Munich, Germany — AgiBot robot official site, 2026-02. https://zhiyuan-robot.com/article/315/detail/139.html
2. AgiBot robot official site homepage (mass production of 15,000 units / double-first at the games / Chimelong theme park / AGILE2.0 / GE-Act 2.0) — AgiBot Innovation, 2026-09. https://www.agibot.com.cn/
3. AgiBot launches the AGILE2.0 perception-control integrated model — AgiBot robot official site, 2026-09-11. https://www.agibot.com.cn/article/315/detail/220.html
4. AgiBot releases GE-Act 2.0, first validation of the native world-action model pretraining and Scaling route — AgiBot robot official site, 2026-09-09. https://www.agibot.com.cn/article/315/detail/218.html
5. AgiBot Southwest Base and data collection center activated — CNR, 2026-05-23. https://www.cnr.cn/scfw/scpd/hygc/20260523/t20260523_527632640.shtml
6. AgiBot robot entry (full chronology through 2026-04) — Baidu Baike, 2026. https://baike.baidu.com/item/%E6%99%BA%E5%85%83%E6%9C%BA%E5%99%A8%E4%BA%BA/63327172
7. AgiBot: leading embodied-intelligence full-stack vendor (in-depth report, including Omdia shipments and data collection factory metrics) — Zheshang Securities, 2026-03-28. https://www.sgpjbg.com/info/df7be3e7242d311ece3785906662f44c.html
8. AI robot weekly report (Unitree IPO / AgiBot's 15,000th unit / IPO process) — NetEase News, 2026-07. https://c.m.163.com/news/a/L1VPHI9505118I96.html
9. IDC "Global Humanoid Robot Market Analysis" (AgiBot and Unitree around 5,000 units metric) — IDC, 2026-01-22 (as quoted in Zheshang Securities and other industry reports).