Embodied AI has been on a hot streak for the past two years.
Robots running, dancing, folding clothes, doing chores — every few months, the industry drops a new demo video. From model capabilities to hardware, embodied AI keeps resetting expectations.
Yet, unlike the buzz in the consumer market, the auto industry is noticeably cooler.
Automakers care about model progress and hardware, sure. But whether a tech actually makes it onto the factory floor depends on engineering verification, not demo reels. Bridging the gap between pulling off a move and keeping pace with the continuous production of hundreds or thousands of cars a day requires an entire manufacturing ecosystem.
On July 3, at the 2026 Embodied AI Industry Scenario Fusion Conference hosted by Gasgoo, a panel titled "What Kind of Embodied AI Partners Do Automakers Need — From Tech Demo to Scale Delivery" brought the issue front and center. Cang Yu, VP and GM of the Industrial Division at Beijing Galaxy General Robotics; Tan Yingcong, Deputy Chief Engineer at Changan Automobile; Fan Jianqiang, Co-founder and COO of FMC³ Robotics; Xu Xiaoshun, Technical Lead for Intelligent Equipment at SAIC-GM Powertrain; and Song Weijia, Director of Final Assembly Process at FAW, gathered to discuss the practical hurdles facing embodied AI in the auto sector.
What Kind of Robots Are Factories Actually Missing?
The first topic: prioritizing robot use cases in the auto industry.
The panelists cited roles that have existed for years with no good fix: material handling, loading and unloading, quality inspection, high-voltage tasks, chassis assembly, wiring harness connection, and parts sorting. These jobs share a common trait — they aren't necessarily complex, but they soak up huge amounts of human labor. Some are physically grueling, others pose health risks, and a few directly impact vehicle quality and production safety.
Logistics is a common entry point. Cang Yu argues that high-load handling, material feeding, sorting, and recycling are highly standardized. It’s easier to form a closed loop there — and more likely to see robots delivered at scale first.
Image source: Xiaomi Robotics
OEMs prioritize quality, safety, and consistency. Tan Yingcong pointed to body gap and flush inspections: a full manual check can take one or two hours, forcing many companies to rely on spot checks. If embodied AI can handle continuous inspection, it would boost efficiency and stabilize results.
Safety and occupational health are another priority. For battery cell assembly or chemical handling, the primary goal is risk reduction. Xu Xiaoshun noted that high-voltage operations and chemical adhesive application offer high value for robots — because companies need to lower risks first, not chase flashy tech.
From the front lines of assembly, the focus narrows to three areas: high-risk safety posts, jobs that strain worker health, and stations critical to quality consistency — such as high-voltage connections, chassis assembly, and wiring harness insertion.
This reflects a clear judgment: new tech shouldn't start with the most complex tasks. Instead, it should fill the gaps where traditional automation falls short and human labor is stretched. On the assembly line, value isn't defined by peak capability, but by the ability to hold down a post reliably, day after day.
From this angle, automakers are more likely to embrace a new tool that solves an old problem. It doesn't need to be stunning — just able to withstand the test of daily repetition.
From Prototype to Production Line: Where’s the Hard Part?
As use cases become clearer, actually getting robots onto the line is far more complex than it looks. For a robotics firm, a prototype nailing a move signals a breakthrough. For an automaker, a successful demo is still miles away from "ready for use."
The two industries define "usable" differently. Robotics chases capability breakthroughs; auto manufacturing demands stability, reliability, and repeatability. A production line runs on a fixed rhythm every day. If one piece of equipment stops, it doesn't just stall one station — it halts the entire line. Automakers need humanoid robots that can maintain precision, rhythm, and quality even after grinding through eight-hour shifts, month after month.
The robotics sector is still in a "developmental phase," according to Cang Yu. Beyond having a general-purpose embodied model, the real test lies in systematic manufacturing and delivery. Auto manufacturing has spent decades building mature final inspection systems. In robotics, many processes still require in-process checks. "When robot production lines move straight to final inspection, that’s when scale delivery truly arrives. It needs both a general-purpose model drive and mature, systematic manufacturing capabilities."
Image source: QianSense Robotics
Once stability is solved, there’s replicability. Making a solution work for one model at one plant doesn’t mean it can be rolled out to others quickly. Different plants have different layouts, vehicle dimensions, and inspection points. For large auto groups, the real question is whether robots can be developed once and adapted fast to different models and factories — rather than redeveloping and recalibrating for every new deployment.
Fan Jianqiang argues that three factors constrain scale: generalization, execution precision, and speed. If every new scenario requires fresh data collection and model retraining, deployment costs won’t drop. His team is using world models, VLA, and atomic skill libraries to build new process primitives, shortening the time robots need to learn new crafts and tasks.
For OEMs, the model is just one piece. Once inside the factory, robots face a battery of engineering issues: solution design, on-site integration, maintenance. Xu Xiaoshun shared his experience deploying humanoid robots this year: from re-development to debugging to ongoing maintenance, every step requires close collaboration between the robot maker and the automaker.
Compared to mature industrial robots, embodied robots are still in the early stages of industrialization. Spare parts systems, repair capabilities, and on-site support all need to be built from scratch. Companies even need contingency plans to prevent a robot failure from derailing the production line.
With auto industry margins under constant pressure, every new investment must prove its worth. Deploying humanoid robots is no exception — the economics have to make sense. Song Weijia revealed that FAW has set up an embodied intelligence subsidiary, aiming to collaborate more deeply with robotics firms on manufacturing scenarios to cut costs, boost quality, and improve efficiency.
It’s clear that scaling embodied AI isn’t just about the tech anymore. Delivery capability, engineering systems, maintenance, cost control, and collaboration between automakers and robotics firms — all of these determine whether robots can actually enter auto factories and run stably over the long haul.
What Determines Whether They Stay?
If engineering capability is the entry ticket, then data and collaboration models determine whether robots get to stay.
Auto manufacturing generates more than just product data. It includes process parameters, production rhythms, quality standards, and decades of accumulated manufacturing know-how. This data touches on core competitiveness — and it’s what OEMs guard most closely.
Cang Yu noted that the industry is currently obsessed with massive datasets to train general foundation models. He suggested that data collection, training, and inference should happen primarily at the edge, supported by continuous pipelines for data and model evolution.
For security reasons, OEMs tend to keep data on-premise, avoiding reliance on cloud training. Tan Yingcong said the industry has been pushing for "data consistency" — ensuring accuracy across different scenarios — then using that scenario-specific data to train and optimize models, rather than just chasing volume. The AI strategy boils down to this: "Define datasets by scenario, then use high-quality datasets to boost model capability."
Fan Jianqiang highlighted data security at the level of national competition. "The last mile of data" must remain in national or corporate hands, much like operators are trained in-house. He argued that the industry needs clear standards for industrial data ownership and IP.
Image source: Luming Robotics
Branding came up repeatedly. Xu Xiaoshun argued that once inside an automaker, robots should act as an extension of brand philosophy, aligning with brand positioning.
Song Weijia sees brand in two layers. The outer layer is what users feel: the Hongqi brand’s intelligent features, styling, and interiors. The inner layer is technical reuse and shared roots — the common ground between autonomous driving, embodied AI, and smart manufacturing. These commonalities can serve as strategic pivots for advancing smart manufacturing and achieving a "second growth curve."
The discussion reflects an underlying attitude in the auto industry: automakers are willing to open their factory floors and try new tech, but collaboration is built on long-term trust. That trust comes from robots doing the job reliably — and from a shared understanding of data boundaries, IP, and how to work together.
For the auto industry, the final deliverable isn't a demo, a model, or a feature — it’s a capability that can participate in manufacturing over the long term.
Where Is This Headed?
As the discussion wound down, a deeper question emerged: When robots take over more stations, how will the manufacturing system itself change? Can we expect to see humanoid robots enter auto factories in volume within the next three to five years?
Cang Yu proposed the concept of "Design for Robot." For decades, process design, logistics flows, and station layouts were built around humans. When AGVs arrived, entire logistics systems were redesigned around their logic. He expects a similar shift driven by embodied AI around 2027.
Tan Yingcong echoed this from an organizational perspective. If robots mature in inspection and assembly, today's human-centric processes and line layouts may need adjustment. The organizational form of "native intelligence" might stop designing workflows around people entirely.
Fan Jianqiang predicts that the first to scale in the next three to five years won’t necessarily be bipedal humanoids. It might be wheeled, dual-arm robots — forms better suited to industrial needs.
In his view, the highly specialized lines of Industry 4.0 have shown their limits in flexibility, hurting economics. The real value of embodied AI is boosting adaptability — letting one line handle more products and processes without constant rebuilding. That shift from specialized to general-purpose is the core evolution from Industry 4.0 to Industry 5.0.
Change is already underway. This year, multimodal large models have handled long-sequence tasks like sorting, grasping, and transport. Xu Xiaoshun noted that in areas like home services, robot cycle times and failure rates are entering an acceptable range, laying the groundwork for scaling up highly standardized tasks.
For the auto industry, though, humanoid adoption will take time. Song Weijia argued that no matter how tech advances, you have to master simple scenarios first: handling, small-part pick-and-place, auxiliary assembly. Only by accumulating data and proving reliability there can you move to complex tasks. That’s the path to faster adoption in the auto sector.
Image source: Wujie Power
Looking back, the panelists came from different corners of the supply chain with different views, but they agreed on one thing: auto factories won’t be torn down and rebuilt because of robots. The real change is gradual integration — robots fitting into the existing system as production methods co-evolve. Shifts in process design, layout, equipment coordination, and management won’t happen all at once. They’ll accumulate project by project.
For most of the panel, the conversation focused on "unsexy" topics that hit closer to reality: delivery, engineering, data, collaboration, and cost.
That "unsexiness" might just be the hallmark of embodied AI’s true industrialization.
The hardest part of introducing tech to manufacturing isn’t the tech itself — it’s whether it can mesh with a mature industrial ecosystem and create value there, consistently and stably.
Tan Yingcong brought up "native intelligence" — the idea that factories and organizations stop designing workflows around humans and instead treat AI capability as a baseline, rethinking line layouts, station setups, and organizational structures. It sounds distant, but it will build up project by project.
Song Weijia had a line that summed it up: "Doing it once beats talking about it ten thousand times." When the industry stops arguing over tech roadmaps and starts debating delivery cycles, failure rates, and spare parts inventory — that’s when embodied AI will have truly crossed the threshold from prototype to production line.
