From Marathon Humanoids to Household Helpers: A Surge in Physical AI
Across the embodied AI ecosystem, activity has shifted from flashy demos to an aggressive race for deployment. A recent humanoid marathon in Yizhuang – where Honor’s “Lightning” shattered the men’s half‑marathon record – highlighted how engineering muscle, not just algorithms, is advancing rapidly, thanks to long legs, liquid cooling and high‑torque motors migrated from consumer electronics design. At the same time, firms are targeting more practical tasks. X Square Robot’s Wall‑B focuses on delicate manipulation in chaotic homes, from picking up trash to arranging flowers, trained on data from more than 100 households to cope with “noisy” real‑world conditions. Compact platforms like XiaoR Geek’s 22‑DOF SamuRoid bring ROS-based humanoids into classrooms and labs, giving developers a low-cost testbed for embodied AI. Together, these efforts show Chinese humanoid robots expanding from spectacle to service, with hardware now robust enough that “the brain hasn’t caught up,” in the words of one founder.
Pudu, Unitree and Agibot: Commercial Leaders in an Embodied AI Stack
Pudu Robotics exemplifies how physical AI is already commercial. The Shenzhen-based company just raised nearly USD 150 million (approx. RM690 million), pushing its valuation above USD 1.5 billion (approx. RM6.9 billion) and cumulative funding past USD 300 million (approx. RM1.38 billion). Its cleaning line now generates over 70% of revenue, industrial delivery robots have shipped more than 4,000 units in their first year, and Pudu claims a 23% global share in commercial service robots. Meanwhile, Unitree’s G1 humanoid is redefining mobility: the wheeled-legged Unitree G1 robot skates, spins and even front-flips on ice and rollers, signalling world-class hybrid locomotion aimed at real-world data collection and task execution. Agibot is pushing a systems view, with its “One Robotic Body, Three Intelligences” architecture and new platforms like the A3 humanoid and G2 Air mobile manipulator, explicitly positioning embodied AI as “productive infrastructure” rather than lab curiosity.
Components and Supply Chains: Lens, Motors and the Hidden Hardware Edge
Beneath the humanoid headlines, a dense robotics supply chain is forming. Lens Technology, best known for consumer device components, now supplies 132 metal structural parts to Honor’s humanoid robot program, illustrating how established precision manufacturers are retooling for embodied intelligence. Analysis of Tesla’s Optimus bill of materials shows why this matters: actuators, reducers, lead screws and motors dominate unit cost, and many of these components are increasingly sourced from Chinese vendors. Each humanoid needs around 28 actuators, 20 coreless motors and 14 reverse planetary roller screws, turning motor and reducer capacity into a strategic chokepoint. By drawing on clustered manufacturing in places like Shenzhen, companies can compress iteration cycles and lower unit costs. Pudu, for example, is explicitly channeling new capital into scaling manufacturing and strengthening its supply chain, building the industrial backbone that separates scalable humanoid deployments from one-off prototypes in Western markets.
The Data Battle: Platforms, Training Grounds and Physical AI Pipelines
If hardware is the body of embodied AI, a new data infrastructure is becoming its nervous system. Tencent’s Tairos platform, Baidu’s embodied intelligence data “supermarket” and JD.com’s trading platform all aim to standardize and monetize physical AI data, from collection to labeling and evaluation. JD plans to mobilize 600,000 people to capture 10 million hours of data and has rolled out a full-stack embodied intelligence data infrastructure spanning acquisition, storage, simulation and testing, positioning itself as a “super supply chain” for robots rather than a robot maker. Home-robot players such as Wall‑B are also emphasizing exposure to messy household environments as critical training ground. Unlike large language models, which feed on closed-loop Internet text, physical AI data must link perception to control signals and outcomes. Companies like Baidu are therefore racing to set the rules for physical AI data flows – a long-term advantage that may be harder for rivals to replicate than any single robot design.
Big Tech, Policy Tailwinds and the Emerging East–West Robotics Divide
Large platform companies are now stitching these pieces into an end-to-end embodied AI ecosystem. Alibaba’s Qwen Micro strategy extends its Qwen model family from dialogue interfaces into physical interaction, with trademarks covering humanoid robots and AIaaS offerings. JD.com is building trading and data platforms while selectively investing in robotics startups. Honor and AutoNavi are transferring strengths from phones and maps into robots, bringing system engineering, heat management and spatio‑temporal data to the field. Policy and state support further accelerate champions like Unitree and Agibot. Compared with Western efforts that often center on high-valuation software or single flag‑ship humanoids, this ecosystem leans on deep manufacturing know‑how, dense motor and reducer supply, and compressed development timelines. With companies already shipping more humanoid robots at lower valuations than some Western peers, the competitive question is no longer whether Chinese humanoid robots can match global standards, but how quickly they will scale and reshape global robotics supply chains.
