Astribot T1 Makes North American Debut at IROS 2026 With Integrated AI Stack and $13K Price

Stardust Intelligence brought its Astribot T1 humanoid to IROS 2026 in Pittsburgh this week for the robot's first North American public exhibition. The pitch is direct: a 1.55-meter, 66-kilogram production robot currently accepting orders at a starting price of approximately $13,000 (89,900 yuan), bundled with a proprietary AI foundation model and an operating system that converts a single-sentence natural-language prompt into a deployable robot application — no low-level control code required.

A Body Designed to Feed the AI, Not the Other Way Around

Astribot calls its design philosophy Design for AI, or DFAI — the premise that mechanical hardware, training-data pipeline, and AI inference architecture should be co-developed rather than assembled from separately optimized components. The clearest expression of that in the T1 is its cable-driven actuation system. Where most humanoid platforms use rigid electric gearbox drives, the T1 routes flexible cable transmissions through its joints, producing natural compliance, lower reflected inertia, and back-drivability. That compliance also means richer proprioceptive training data — force and position signals with physically realistic continuous variation rather than the step-function behavior typical of rigid drives. The T1 has 23 degrees of freedom and a five-kilogram payload per arm.

How Lumo-2 Replaces Text-Based Planning With Latent World Prediction

The foundation model on T1 is Lumo-2, described in an arXiv paper submitted July 13, 2026 by the Astribot team at Stardust Intelligence. Lumo-2 is a Latent World-Action Model — a meaningful architectural departure from the first generation of language-model-guided robot control.

Earlier systems, including Astribot's own Lumo-1, used explicit textual chain-of-thought as a planning layer: the model verbalized its intended actions before converting them into motor commands. Lumo-2 eliminates the text-format intermediate step. A Vision-Language Encoder processes incoming sensory data; a Latent World Dynamics Model predicts future world states as physically grounded latent representations of visual transitions — an internal simulation of how the scene is about to change; an Action Generator converts those predictions into motor commands. Planning happens in latent space, not in words. The company argues this produces faster inference and better handling of continuous real-world dynamics than text-chain approaches.

This places Lumo-2 in the same conceptual territory as world-model approaches from model-based reinforcement learning — systems that maintain an internal environment model rather than mapping directly from observation to action.

Read more: World Action Agent: VLMs Rehearse Robot Moves Before Physical Execution

The Pricing Gap Astribot Is Targeting

The humanoid developer platform market is sparse enough that pricing carries outsized signal. Unitree's G1 has established a low-end entry point at approximately $13,500, with limited developer tooling at that base price. Enterprise platforms such as Agility Robotics' Digit target large commercial deployments at substantially higher price points and are typically delivered through structured commercial programs. Astribot is placing the T1 at approximately $13,000 with stated immediate delivery, targeting the underserved middle of that market — developers who need more than a bare hardware frame but cannot justify enterprise pricing.

Read more: Agility Robotics Explores Wheeled Robots as Digit 5 Launches and SPAC Deal Nears

AOS and the Case for a Robot App Store

The software layer may carry the longer-term competitive implication. Astribot OS provides a layered SDK — joint-level control, Cartesian-space planning, whole-body coordination, sensor abstraction — with a natural-language compiler on top that converts a single human sentence into a packaged, schedulable robot application. OTA updates deliver model improvements to deployed units without field retraining.

At a developer hackathon held September 21 in Beijing, fifteen teams used AOS to build more than ten functional robot applications in 36 hours from single-sentence prompts, according to the company. Those results are company-reported and come from a controlled environment, but the throughput suggests the abstraction level is functional. If robot skills become packageable and distributable — the "robot app store" Astribot has described — the value chain shifts toward software and away from hardware, which is a different competitive moat than most humanoid platforms are building toward.

Limitations and Open Questions

Several significant questions remain unresolved. Lumo-2 is described in a company-authored arXiv paper without independent peer review. No neutral third-party benchmark comparison of T1 against Unitree G1 or other platforms has been published. The hackathon results are self-reported. Immediate delivery and the stated starting price are company claims that independent buyers have not publicly confirmed.

Critically, Lumo-2 does not learn in the field. Capability improvements require a company-managed OTA push. A deployed T1 will not improve from its own experience between updates — a real constraint for teams that expected continual on-device adaptation. Data residency terms for the AOS cloud connection have not been publicly detailed.

What Comes Next

IROS 2026 closes October 1. The AOS Hackathon Finals are scheduled for October 24–25 in Shenzhen, where teams from the September event will present refined applications. That competitive context will offer a more structured read on what developers have actually built on the platform.

The more consequential question is whether the latent world-action architecture generalizes outside controlled demos at a price point that lets enough developers test it. Physical AI has consistently required more deployment iteration than laboratory results predict. The T1's approximately $13,000 starting price and stated immediate availability at least lower the barrier to finding out.