At IROS 2026 in Pittsburgh, Astribot put a specific proposition before a North American robotics audience: a humanoid product built around a continuously updated AI-model layer, an embodied operating system, and a cable-driven body. Robotics & Automation News reported that the Shenzhen company showcased Astribot T1 in North America for the first time at the event. The publication dates the Pittsburgh conference to September 27 through October 1 and identifies the David L. Lawrence Convention Center as its venue.
Astribot identifies the business in English as Astribot Co., Ltd., says it was founded in Shenzhen in December 2022, and names the product Astribot T1. The Pittsburgh showing centered on the company’s integrated physical-AI approach rather than a single hardware feature. That framing matters because the reported T1 presentation connected its model family, operating software, mechanical architecture, demonstrations, developer interfaces, and commercial availability claims in one package.
Lumo Models Anchor Astribot’s Integrated Physical-AI Architecture
Robotics & Automation News describes the Lumo model family as the AI component of that approach. According to the publication, Lumo-1 introduced what Astribot calls a Reasoning-Action Foundation Model approach. Lumo-2, which the company says is T1’s base model, uses a latent connection between world representations and action for harder physical tasks. Those descriptions identify the company’s intended model-development path; they do not establish independently validated task performance, reliability, or comparative benchmark results.
The same coverage places the model work alongside two other elements: Astribot’s embodied operating system and its cable-driven physical structure. Astribot characterizes the combination as an integrated architecture, with the model layer, software layer, and physical structure designed to work together. The company’s Design for AI approach is therefore a company-described design framework tied to the T1 presentation, not evidence of a completed deployment at scale. The Pittsburgh appearance supplied a public setting for Astribot to show how it wants those layers to be understood together.
The training and update cycle described for this stack also sets a clear boundary around the claims. The publication says Astribot collects and curates data, retrains its models, and later deploys updated skills. According to the account, a deployed T1 does not autonomously learn as it works in real time. That distinction is material for interpreting the reported capabilities: an observed task demonstration reflects the model and skill version presented, while subsequent changes depend on the company’s stated collection, curation, retraining, and deployment process.
T1 Demonstration Pairs Deformable-Object Handling With Documented Specifications
The reported T1 demonstration involved autonomously tidying miscellaneous items into a backpack, including the manipulation of deformable objects. Astribot and Robotics & Automation News present that sequence as a demonstration of the Lumo-2-equipped machine. It supports a narrow observation about what was shown, rather than a claim that T1 can handle every household, industrial, or service task with the same result. No independent safety evaluation, customer-deployment evidence, or performance benchmark is identified in the cited materials.
The reported specifications put the demonstration in physical terms. T1 is approximately 1.55 meters tall and weighs about 66 kilograms, according to Robotics & Automation News. The platform is listed with 23 degrees of freedom before end effectors, plus a per-arm payload rating of 5 kilograms. Those figures describe the product presented by Astribot; the coverage does not supply independent testing of the measurements or payload performance. T1’s profile adds a product-level thread to EastFrontier’s recent coverage of UBTech opening a 10,000-unit humanoid robot factory in Liuzhou, while the two developments remain separate company events.
Astribot’s reported hardware design also includes automatic charging and quick battery replacement. The report says the platform is modular across end effectors, computing modules, and sensors. Together with the 23-degree-of-freedom figure and per-arm payload claim, those stated design choices explain why the company presented T1 as a platform with adaptable hardware elements. They do not reveal battery runtime, operating speed, uptime, task accuracy, or a validated operating envelope, none of which is provided by the available materials.
SDK Interfaces Define the Next T1 Questions
For developers, the report says T1 provides SDK and API access covering joint control, Cartesian motion, whole-body coordination, and sensor data. These interfaces are part of the product description, and Astribot’s presentation does not disclose which applications North American users have built with them or how third parties have evaluated the interfaces. The developer-access emphasis also fits the wider physical-AI activity tracked by EastFrontier, including Renesas opening a Beijing lab for physical AI and robotics, a separate development identified in the site’s archive.
The commercial language requires equally close qualification. Robotics & Automation News says U.S. customers can order the robot and that immediate delivery is available. These are reported Astribot commercial statements. They do not confirm inventory levels, shipment volumes, completed sales, regulatory clearance, or customer uptake.
Astribot’s first reported North American T1 showcase at IROS 2026 consequently provides a defined snapshot: a Pittsburgh appearance, a Lumo-2-based humanoid, a company-described model-to-body architecture, a backpack-tidying demonstration, published physical specifications, and developer interfaces. The available information also fixes the limits of that snapshot. The appearance does not establish a broader overseas launch, mass deployment, U.S. sales volume, or independently validated performance. What is established is the company’s decision to present T1’s integrated physical-AI architecture and reported commercial availability before the IROS audience in Pittsburgh.