Chinese Embodied AI Startups Draw 93.5 Billion Yuan in First Half
In this article (4)
Chinese embodied artificial intelligence developers raised 93.5 billion yuan ($13 billion) across 322 funding deals in the first half of 2026, according to data research firm ITjuzi. That figure represents a fivefold jump from the same period a year earlier, shifting capital from mechanical hardware into the software models that direct autonomous machines.
Hardware makers are slashing retail prices on bipedal frames and robotic arms to secure market share, destroying manufacturing margins in the process. The shift forces commercial value into cognitive models capable of handling physical labor without custom engineering for every task.
Where the Capital Lands
Three specialist software developers, X Square, Galaxea, and Spirit AI, have each reached private valuations of roughly $3 billion, based on company disclosures. Hong Kong-listed UBTech advertised an executive compensation package of up to 124 million yuan ($18 million), including equity and benefits, to recruit a chief scientist for its AI model team. Hangzhou-based Unitree, one of the world’s largest humanoid robot producers by volume, now maintains a team of roughly 30 researchers dedicated to spatial physics and predictive object behavior.
Hardware has become a commodity faster than expected across supply chains. Chinese builders can assemble a dynamic bipedal chassis at a fraction of Western production costs, but those machines remain commercially useless if every warehouse shelf or retail counter requires months of custom code. The venture money is now chasing generalist models that allow an off-the-shelf unit to unload pallets, stock pharmacy displays, or thread industrial wiring with minimal training.
That dynamic alters procurement calculus for logistics networks, retail operators, and assembly plants across Asia. Instead of paying hefty integration retainers to robotics vendors, supply chain directors are testing software subscriptions that deploy across mixed fleets of machines. The operational risk lies in real-world error rates: while simulated environments and human video clips allow rapid laboratory demonstrations, factory floors and convenience stores tolerate zero mistakes.
Data Collection in Pharmacies and Plants
Different developers are testing competing training methodologies to feed physical data to their networks. Spirit AI hired roughly 1,000 workers to record routine manual tasks and industrial activities, according to co-founder Gao Yang. ACE Robotics strapped wearable sensors onto store clerks and hotel hospitality staff to record precise grip strength and hand movements during service shifts.
“Spirit AI hired roughly 1,000 workers to record routine manual tasks and industrial activities, according to co-founder Gao Yang.”
Ant Group’s robotics unit, Robbyant, trained its model using 160,000 hours of operating material, which included 50,000 hours of human-guided mechanical movements. The group tested its software across 20 robot configurations built by 17 different equipment brands, seeking a cross-platform operating system for physical automation. In Shanghai, Robbyant deployed robots into two commercial pharmacies to pick and organize retail prescriptions, leaving human pharmacists to inspect the items before dispatch.
Industrial plants demand higher precision than retail counters. Battery manufacturer CATL integrated Spirit AI machines into its Zhongzhou production facility, running flexible wire assembly during pack testing. The battery maker reported an insertion success rate above 99 per cent on that specific line.
Simulations Meet Hardware Bottlenecks
Other software houses bypass physical recordings entirely. DexForce uses synthetic computer-generated training environments to prepare robotic arms for precision tasks within six to eight hours, reporting operational success rates above 99.9 per cent on defined factory assignments. However, spatial intelligence developers such as Manycore Tech caution that computer graphics fail to simulate micro-friction, material deformities, and changing warehouse lighting conditions.
The push into advanced physical modeling faces immediate structural constraints inside mainland laboratories. United States export controls on advanced Nvidia processors limit access to the compute clusters needed to train large vision-language-action models. Chinese startups are turning to synthetic data pipelines, model pruning, and domestic processors developed by Huawei to bypass the compute bottleneck, but the processing gap widens as networks scale.
Next Steps for Industrial Deployment
The transition from bespoke factory automations to adaptive common-sense robotics marks the next phase of China’s robotic push. Unitree founder Wang Xingxing reported experimental success testing OpenAI’s GPT-6 model on walking units, setting an operational target where robots complete 80 per cent of assignments across 80 per cent of unfamiliar retail and factory environments via direct voice instructions.
Whether domestic models can match that benchmark without unrestricted access to foreign silicon remains the central variable for Chinese supply chain tech, and will show whether these 93.5 billion yuan in capital commitments translate into profitable labor replacement.
Reader pulse
Can generalist AI models make retail humanoid robots commercially viable?
21,537 votes so far