The pin holes are so small you have to squint to align them, and even when you think you've got it, you might have mis-inserted — automotive wiring harness assembly is called the “deep water” of industrial automation; workers train for a month, and many don't last three months on the job before quitting.

At WAIC 2026, embodied AI company Tashihang brought a 1:1 scale circular wiring harness assembly line directly onto the exhibit floor. Multiple robots collaborated in clusters; the line rotated harness boards; each station assembled different harness types simultaneously. Tashihang Chief Scientist Ding Wenchao even joked: “On site we don't dare let it run continuously — we're afraid the robots will work too fast and run out of materials.” Powering this line was the day's newly released embodied native foundation model AWE 3.5, which also took home WAIC's highest honor, the SAIL Star award — the only general-purpose embodied large model to win the award this year.

The Native Model Approach: Rebuilding the Embodied Brain from First Principles

This year the embodied AI community has been debating whether VLA (Vision-Language-Action) or world models represent the future, but people actually training models don't care about these labels — they care about one thing: does it optimize Action?

The VLA architecture's problem is that vision and language are tightly coupled from pretraining, but the action modality isn't connected until post-training via SFT. It's the equivalent of a student learning all the theory first, then starting lab work the night before the exam. The world model approach leverages large-scale video for representation learning, but internet video models' attention structures have already been shaped by millions of hours of cat and dog videos; plugging these directly into an embodied foundation causes two issues: video model hallucinations leak into action policies, and visual understanding never synchronizes with action.

Tashihang chose to start from scratch — training a natively embodied model from the ground up. From pretraining, AWE 3.5 feeds vision, language, action, and other modalities into the model together; the same architecture can both output actions and predict the future. Ding calls this the “One Model structure”: the model itself doesn't change; only the input-output configuration changes to switch between “personalities.” Vision-to-action is the Base Policy; action-conditioned future visual prediction is the Action-Conditioned World Model; their interaction forms a complete reinforcement learning loop.

Embodied Scaling Has Fully Arrived

AWE 3.5's core breakthrough: the model becomes its own simulator.

Traditional robot training requires trial and error on real hardware — zip a zipper, pull crookedly, try again; wrong force gets it stuck, try again. Data is expensive, speeds are slow, hardware wears out. AWE 3.5's approach is to first simulate inside the model's “mind,” imagining outcomes under different forces, then feeding the optimal policy to the Base Policy — faster than real-world trial-and-error by more than an order of magnitude. Pretraining establishes general physical interaction capability; post-training enables self-play and self-reinforcement inside the model's internal world.

What makes this paradigm work is the convergence of two hidden threads: over a million hours of human-centric real-world data, and the infra to digest it. AWE 3.5 is trained on over a million hours of human-centric data — a scale comparable to what video generation models and autonomous driving require. Crucially, new tasks now need only hours of data collection to reach a roughly usable level. That means Scaling's power has been validated in embodied AI.

Ding offered a clear assessment: embodied scaling is fully unlocked, and a watershed moment likely arrives between H1 and mid-2027. The industry has moved from “robots dancing for show” to “robots actually working.” Over 200 embodied AI companies exhibited at WAIC this year, with the robotics area drawing noticeably larger crowds than other zones — exhibit traffic is a leading indicator. After AI Coding forced LLMs to build data flywheels and Harness Engineering systems, embodied AI now awaits its own “killer app” closed-loop scenario. Wiring harness assembly might just be the first.

See you tomorrow.