AgiBot

Scroll sideways for more →

AgiBot, styled AGIBOT and branded in Chinese as Zhiyuan Robotics (智元机器人), is a Chinese robotics company founded in Shanghai in February 2023. It develops robot hardware, data-collection systems, simulation software, and robot-learning models. Its hardware portfolio spans full-size and compact humanoid robots, wheeled dual-arm task robots, quadruped robots, commercial cleaning machines, and dexterous hands.^[1][2]

The company said its 15,000th robot rolled off the production line on June 28, 2026. That figure is a cumulative factory milestone across AgiBot's portfolio, not a disclosed count of paid deliveries or active deployments. Separately, Omdia's 2026 market radar lists 5,168 AgiBot shipments for 2025 and a 39 percent share of the report's general-purpose embodied-intelligent-robot segment. Omdia's vendor analysis explicitly describes the shipment number as AgiBot's claim, and the report's category includes wheeled as well as bipedal forms.^[3][4]

This article describes the company as it stood at the editorial cutoff of July 28, 2026, 23:59:59 in Bangkok (UTC+7).

History and leadership

AgiBot was established in February 2023 and introduced its first full-size bipedal platform, the RAISE A1, that August. It opened a manufacturing facility in Shanghai in January 2024. In August 2024 it presented the A2, A2-W, A2-Max, X1, and X1-W, with commercial deliveries planned from the Lin-gang facility. The company opened its AIDEA data-collection factory in September 2024 and reported the production of its 1,000th general-purpose embodied robot in January 2025.^[1][2]

The next factory announcements came at shorter intervals. AgiBot reported 5,000 cumulative units in December 2025, 10,000 in March 2026, and 15,000 in June 2026. These are manufacturer announcements and use changing labels, so they are best read as production checkpoints rather than a continuous audited series.^[3][5][6]

Scroll sideways for more →

Deng Taihua was identified by the company in May 2026 as founder, chairman, and CEO. Peng Zhihui was identified as co-founder, president, and CTO. Peng previously worked at Huawei and became known for engineering projects published under the name Zhihui Jun. His role is technical and executive; calling him the company's sole founder would omit Deng and the wider founding team.^[7][8]

Products

AgiBot groups much of its portfolio into A, X, and G families, while also selling D-series quadrupeds, the C5 cleaning robot, and OmniHand end effectors. The word "humanoid" is used broadly in company material and market reports: some G-series machines have a humanlike upper body on an omnidirectional wheeled base rather than legs.^[4][9]

Scroll sideways for more →

At the World Artificial Intelligence Conference on July 18, 2026, AgiBot announced four additions: the A3 Ultra full-size humanoid, X2 Edu research platform, G2 Max heavy-payload wheeled task robot, and OmniHand 3 Ultra-M direct-drive dexterous hand. The announcement establishes that these products had been unveiled by the cutoff, but it does not by itself establish their delivery volume, field reliability, or long-term availability.^[10]

Manufacturing and deployment evidence

Production, shipment, delivery, and deployment measure different things:

Scroll sideways for more →

The clearest disclosed industrial case at the cutoff was AgiBot's work with electronics manufacturer Longcheer Technology. In April 2026, the companies said multiple G2 robots had been integrated at multimedia integrated testing stations on an active tablet line. The robots loaded and unloaded tablets, placed them in test fixtures, and sorted results. A later six-day livestream showed multiple robots on the line. AgiBot published aggregate task and output figures after the broadcast, but those figures remain supplier and customer reports rather than third-party operational studies.^[11][12]

The distinction matters because the same robot can be counted once when produced, once when shipped to an integrator, and again in a deployment announcement. None of the cited sources provides a serial-number-level reconciliation among those stages. Omdia nevertheless placed AgiBot, Unitree, and UBTech in its first tier of vendors with more than 1,000 shipments in 2025, and independent reporting described AgiBot and Unitree as the only two above 5,000 in that dataset.^[4][13]

Data, models, and software

AgiBot World and GO-1

AgiBot World is a real-robot manipulation dataset and training platform developed with OpenDriveLab, the University of Hong Kong, Shanghai Innovation Institute, and Shanghai AI Laboratory. The 2025 paper reports 1,001,552 trajectories totaling 2,976.4 hours, covering 217 tasks, 87 skills, and 106 scenes in five domains. More than 100 AgiBot G1 robots collected demonstrations through teleoperation, with manual review and failure-recovery examples included in the pipeline.^[14]

The public repository later labeled its complete Beta release as 1,003,672 trajectories and its curated Alpha subset as 92,214 trajectories. The difference from the paper is a version difference, not a reason to merge the two counts. The repository releases its data and code under CC BY-NC-SA 4.0, which permits reuse under attribution and share-alike conditions but excludes commercial use.^[15]

GO-1 is the vision-language-action policy introduced with AgiBot World. Its public model card describes a 3-billion-parameter model based on InternVL 2.5-2B and uses the same CC BY-NC-SA 4.0 license. The paper's reported gains come from author-run experiments on its specified robots, datasets, and tasks; they are not a certification of general robot competence.^[14][16]

World models, simulation, and post-training

AgiBot's later research expanded from offline demonstrations toward prediction, simulation, and learning from deployed fleets.

Scroll sideways for more →

Genie Envisioner was released as a preprint in 2025. Genie Sim 3.0, SOP, and Learning While Deploying followed in 2026. ACoT-VLA is the strongest peer-reviewed evidence among the named GO-2 components at the cutoff because it appeared in the CVPR 2026 proceedings. The larger GO-2 product combines additional company-described elements, so conference acceptance of ACoT-VLA should not be extended to every GO-2 claim.^{[17][18][19][20][21]}

The May 2026 tau0-WM report describes a 5-billion-parameter video backbone plus an action component trained on about 27,300 hours of robot teleoperation, UMI-style data, and egocentric human video. Heterogeneous sources supervise different outputs: human video supplies visual dynamics but not robot-control labels. This is an important scope boundary because hours of video are not interchangeable with hours of executable robot demonstrations.^[22]

Open-source scope

AgiBot publishes selected research, software, and hardware resources, but there is no single license covering the whole company stack:

• AgiBot World data, repository code, and GO-1 weights use CC BY-NC-SA 4.0.^[15][16]
• The ACoT-VLA implementation uses Apache 2.0.^[24]
• Major Genie Sim source modules use Mozilla Public License 2.0, while its scene-reconstruction directory and third-party components carry multiple licenses.^[25]
• AimRT, a C++20 runtime developed and used by AgiBot, is published under Mulan PSL v2 and documents compatibility with ROS 2, HTTP, and gRPC ecosystems.^[23][31]

"Open source" therefore applies to named artifacts, not automatically to commercial robot firmware, training data, cloud services, third-party dependencies, or every model discussed by the company. Users need to check the license file and revision for each repository or model they adopt.

Corporate status

Reuters reported in August 2025 that AgiBot completed a strategic financing round involving LG Electronics and Mirae Asset. Reuters also reported in October 2025, citing unnamed sources, that AgiBot was considering a Hong Kong initial public offering and had appointed banks to work on it. Those plans and valuation targets were reported proposals, not completed transactions.^[26][27]

On July 24, 2026, Reuters reported that the state-run Securities Times said AgiBot had initiated the Hong Kong IPO process. The report establishes a reported process step, not a completed filing or listing. As of the article cutoff, the earlier valuation range should not be presented as the company's current market value.^[28]

Evaluation, safety, and support limits

Several evidence boundaries recur across AgiBot's public record:

• Manufacturer metrics: Production checkpoints, Longcheer task counts, and most product performance figures were published by AgiBot or a deployment partner.
• Market estimates: Omdia's report is independent market analysis, but it labels AgiBot's 5,168-unit shipment figure as the company's claim and groups bipedal and wheeled robots in one segment.^[4]
• Research generalization: The research papers use defined datasets, platforms, tasks, and success criteria. Genie Envisioner states that its training corpus was single-platform and its experiments did not cover full-body locomotion or dexterous-hand coordination. Learning While Deploying says its policy-learning framework did not explicitly model execution safety.^[17][21]
• Simulation limits: Genie Sim reports large-scale synthetic data and controlled sim-to-real experiments. Simulation coverage does not establish reliability in every physical environment.^[18]
• Commercial evidence: No cited source provides audited fleet uptime, maintenance cost, customer retention, injury rates, or a serial-number-level map from production to active deployment.

AgiBot published a vulnerability-disclosure policy directing security reports to security@agibot.com. It says the company will respond within five working days and aims to triage within ten, while prohibiting destructive testing and denial-of-service activity.^[29]

The company's defined-support-period document promises security updates for at least one year from the listed launch date for covered devices. At the cutoff it listed January 1, 2027 as the end of security-update support for the A2, X2 Ultra, and G2, and the same date for several D1 and X2 Neo configurations. It says published periods will not be shortened and may be extended. Operators should verify the current document for each exact model rather than infer software support from a product's continued sale or physical service life.^[30]

References

1. ^AgiBot, "About Us", accessed July 2026.
2. ^Shanghai Municipal Government, "Shanghai's first humanoid robot factory prepares for delivery", August 20, 2024.
3. ^AgiBot, "AGIBOT's 15,000th Robot Rolls Off the Production Line, Marking a New Milestone in Embodied AI Deployment", June 28, 2026.
4. ^Omdia, *Market Radar: General-purpose Embodied Intelligent Robots, 2026*, January 2026.
5. ^AgiBot, "AGIBOT Announces the Rollout of Its 5,000th Mass-Produced Humanoid Robot", December 8, 2025.
6. ^AgiBot, "AGIBOT Reaches 10,000 Units as Real-World Demand for Robots Accelerates", March 30, 2026.
7. ^AgiBot, "The First Hong Kong Embodied AI Industry Summit and AGIBOT Partner Conference 2026 Hong Kong Opens", May 12, 2026.
8. ^Xinhua, "Profile: A Chinese prodigy on a mission to bring robots to life", May 3, 2026.
9. ^AgiBot product catalog, accessed July 2026.
10. ^AgiBot, "AGIBOT Unveils Four New Products at WAIC 2026, Showcasing Embodied AI in Real-World Operations", July 18, 2026.
11. ^AgiBot, "AGIBOT and Longcheer Technology Achieve World's First Embodied AI Deployment in Consumer Electronics Precision Manufacturing Mass-Production Line", April 14, 2026.
12. ^AgiBot, "AGIBOT Concludes Six-Day Real Factory Livestream, Signaling Shift from Humanoid Robot Demonstrations to Deployment-Led Validation", June 29, 2026.
13. ^Associated Press, "Humanoid robots show off their language and boxing skills in Hong Kong", April 13, 2026.
14. ^AgiBot World contributors et al., "AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems", arXiv:2503.06669, revised August 4, 2025.
15. ^OpenDriveLab AgiBot World repository, accessed July 2026.
16. ^AgiBot World GO-1 model card, accessed July 2026.
17. ^Liao et al., "Genie Envisioner: A Unified World Foundation Platform for Robotic Manipulation", arXiv:2508.05635, revised November 4, 2025.
18. ^Yin et al., "Genie Sim 3.0: A High-Fidelity Comprehensive Simulation Platform for Humanoid Robot", arXiv:2601.02078, revised June 30, 2026.
19. ^Zhong et al., "ACoT-VLA: Action Chain-of-Thought for Vision-Language-Action Models", *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, 2026, pp. 8152-8162.
20. ^Pan et al., "SOP: A Scalable Online Post-Training System for Vision-Language-Action Models", arXiv:2601.03044, January 6, 2026.
21. ^Wang et al., "Learning While Deploying: Fleet-Scale Reinforcement Learning for Generalist Robot Policies", arXiv:2605.00416, revised June 3, 2026.
22. ^AgiBot Finch, "tau0-WM: A Unified Video-Action World Model for Robotic Manipulation", May 31, 2026.
23. ^AimRT repository and documentation, accessed July 2026.
24. ^AgiBotTech ACoT-VLA repository, accessed July 2026.
25. ^AgiBotTech Genie Sim repository, accessed July 2026.
26. ^Reuters, "Chinese robot maker AGIBot completes new round of financing", August 1, 2025.
27. ^Reuters, "Chinese robot maker AgiBot plans Hong Kong IPO next year, sources say", October 10, 2025.
28. ^Reuters, "Chinese robot maker AgiBot starts Hong Kong IPO process, Securities Times reports", July 24, 2026.
29. ^AgiBot, "Publication of Vulnerability Disclosure Policy", May 11, 2026.
30. ^AgiBot, "Defined Support Period", May 21, 2026.
31. ^AgiBot, "AgiBot open-sources the AimRT robot middleware", September 25, 2024 (Chinese).