AI is getting Physical — and China knows it
Embodied AI and moving beyond the brain in the cloud
When OpenAI’s GPT-2 began generating coherent paragraphs in 2019, it felt like a remarkable breakthrough. By 2023, ChatGPT was composing poetry, drafting legal memos, and writing startup code. Silicon Valley collectively wondered: had we finally reached general intelligence?
Yet, despite these impressive feats, this kind of AI remains fundamentally limited. It cannot perform simple physical tasks—screwing in a lightbulb, flipping a dosa, or picking up a crumpled napkin without breaking it. This limitation is not a minor flaw but a profound blind spot in how intelligence is commonly defined. We have conflated language with cognition and text prediction with genuine reasoning. Most of today’s “smart” systems are like brains trapped in jars—disembodied, disconnected, and essentially useless in the physical world where real complexity unfolds.
Meanwhile, a quieter revolution is taking shape. From Shenzhen to Seoul, and increasingly across the Global South, the next phase of AI focuses not on what machines can say, but on what they can do.
This AI is embodied, embedded, tactile, and deeply intertwined with the messy, unpredictable reality around us.
The Problem with Brains in Jars (Vats)
This philosophical divide is not new. As early as 1997, philosopher Andy Clark argued that intelligence does not reside solely in the brain but emerges from a continuous feedback loop among mind, body, and environment. A toddler doesn’t learn gravity by reading about it; she learns it by dropping the same cup dozens of times.
Yet, today’s AI labs treat cognition as a software problem: train a model on enough data, and it will understand the world. In reality, it won’t. AI can describe a bicycle perfectly but has no clue how to ride one. It can write about cooking techniques but will burn toast every time.
Robotics pioneer Rodney Brooks put it plainly: “The world is its own best model.” You don’t simulate walking—you walk, stumble, learn, and adjust.
Two Diverging Paths in AI Development
This divide is playing out in real time. The West, led by the U.S. and Europe, has focused heavily on large language models and software-native intelligence—cloud-based oracles that live in browser tabs and sound increasingly sentient.
Across Asia, especially in China and South Korea, the focus is shifting toward AI that moves: delivery robots, assembly-line arms, hospital assistants, and even humble appliances like rice cookers that adapt to different grain types.
For example, in Shenzhen, a ¥399 smart rice cooker uses computer vision to identify the type of rice and automatically adjusts water ratios. It may not sound like AI, but it’s quietly effective and very much alive in the physical world. Meanwhile, in the West, $150,000 robotic kitchens still struggle with basic tasks like dicing an onion.
The point is not to idealize China (it has flaws!) but to highlight how embodiment changes the game. Countries innovating in the physical realm have a head start on what’s coming next.
Where AI Is Becoming Physical
1. Learning on the Factory Floor
At Foxconn’s advanced plants in Shenzhen, robotic arms equipped with real-time vision and reinforcement learning detect defects, adjust their grip in milliseconds, and optimize workflows on the fly. These systems learn continuously during their shifts, not waiting for quarterly software updates.
While Western companies like Boston Dynamics and Covariant are pushing boundaries, deployment is slower. Pilots often stall in demo phases or get bogged down by hardware procurement, mechanical challenges and safety reviews. By the time approval comes, production lines have moved on.
2. Bots That Navigate Crowded Sidewalks
Meituan’s delivery robots in Shanghai have completed over 10,000 deliveries, handling more than 30 deliveries daily. They don’t rely solely on GPS but use simultaneous localization and mapping (SLAM), real-time processing, and swarm coordination to dodge electric scooters, toddlers, and parked cars.
In contrast, similar robots in the U.S., like those from Starship Technologies, remain confined to gated areas such as college campuses and business parks. This is not due to technological failure, but because the infrastructure and social norms don’t support their operation in public spaces.
This highlights the concept of “embodied infrastructure”: the roads, regulations, and cultural expectations that allow physical AI to function in the real world.
Video courtesy: Meituan
3. Hospitals That Anticipate Needs
At places like Tianjin Medical University, robots don’t just assist—they anticipate. Surgical robots connected over 5G monitor patient vitals in real time, predicting complications before they arise. Autonomous UV disinfecting robots reduce infection rates by up to 30%.
Western hospital robotics, still dominated by high-precision systems like the da Vinci robot, remain largely reliant on human teleoperation and lack predictive or autonomous feedback loops.
But not every application needs to be surgical.
The real low-hanging fruit lies in simpler, high-frequency tasks—like autonomous robots delivering medications, assisting with post-operative care, or supporting non-critical in-patient monitoring. These are not moonshots—they're practical, scalable interventions that could free up staff time and improve consistency in care.
The key question is not who is ahead technologically, but who adopts and integrates these systems faster—and what happens when AI becomes predictive and embedded rather than merely reactive.
Picture courtesy: Shanghai Children’s hospital
Why This Shift Is Happening Elsewhere First
1. Policy as Infrastructure
Unlike the U.S., which largely leaves AI innovation to private companies, China’s state-led strategy actively supports embodied AI deployment, subsidizing robots in hospitals, logistics, and hospitality. This accelerates feedback loops: cities build infrastructure to test autonomous taxis; hotels receive funding for cleaning robots.
This approach is not inherently better as it involves trade-offs in regulation, privacy, and risk tolerance, but it does enable faster progress.
2. Ubiquitous 5G and Edge Computing
China boasts 3.5 million 5G base stations compared to about 500,000 in the U.S. This vast network enables local processing of sensor data, critical in environments like busy kitchens or surgical theaters where latency matters. Edge AI and the ability for machines to think and act in real time, near the data source is the nervous system enabling embodied intelligence.
3. A Culture of Trial and Error
In China, rapid tech adoption and a high tolerance for experimentation mean innovations like facial recognition payments or autonomous room service rollouts rarely spark regulatory panic. While this raises valid concerns about surveillance and labor displacement, it also creates momentum. AI improves through deployment, not just research.
What Comes Next: Cobots (Collaborative Robots), Not Chatbots
The future is unlikely to be machines replacing humans but rather machines becoming collaborators among themselves and alongside humans in factories, kitchens, hospitals, and homes.
Robotic arms like Flexiv’s Rizon learn tasks by sensing resistance.
Hospital cobots flag anomalies in patient vitals during surgery.
Domestic robots grow more context-aware like your Roborock (China’s answer to the Roomba) knows your floor plan; soon, or, your fridge will know when you’re running low on eggs.
Embodied AI is less about consciousness and more about competence. It’s not here to chat; it’s here to help.
Why the Body Still Matters
The West is building brilliant minds in the cloud with large language models that are fluent and persuasive. But without a body, they remain what philosopher David Chalmers might call “zombies” or simulations of understanding with no real contact with the world they describe.
Asia, despite its flaws, is betting on something more grounded and the physical/hardware. Not necessarily better, but different. AI that acts, listens, learns through motion, stumbles, adjusts, and tries again.
The gap between these approaches is growing—not in raw capability, but in context. The sooner we bring AI out of the cloud and into the world, the sooner it can start doing something truly useful.
The real question is not whether AI is smart. It’s whether it is embodied enough to matter.