While the entire embodied AI industry discusses entering factories and retail stores, Geekplus (02590) co-founder and CEO of Geekplus Embodied AI Technology Co., Ltd., Chen Xi, offers a starkly different perspective. His logic strikes at the core: manufacturing scenarios have poor replicability and are difficult to scale. Every workstation on every production line in every industry is different. Getting embodied robots into factories requires extensive manual data collection and deployment, coupled with the high cost of humanoid robots, making the overall economic viability uncompetitive. For retail scenarios like convenience stores and pharmacies, "unmanned retail can be achieved without robots; embodied entry is not a real pain point or need." The retail demonstrations seen at exhibitions have neatly arranged products, but in real stores, cluttered, stacked, and obscured items are the norm. Precisely picking a single item from a pile of goods is an order of magnitude more difficult.
In contrast, warehousing is the ideal training ground and commercial starting point for embodied AI. Here, operational standards are highly unified, but SKU counts are enormous and data density is extremely high. This environment best cultivates embodied technology capabilities, achieves commercial closure, and enables large-scale replication. At the WAIC 2026 exhibition, Geekplus delivered on this judgment. The humanoid robot Gino 1 collaborated with a mobile robot fleet, seamlessly completing picking, box-carrying, and delivery tasks in a micro-warehouse. There were no slow-motion sequences or remote-controlled shots; it was the only intelligent productivity team on site ready for immediate deployment.
Efficiency is 1, Generalization is the 0 Behind It
When the industry widely regards generalization capability as the highest benchmark for embodied AI, Chen Xi offers a sharper viewpoint: "Efficiency is the ticket to commercialization. Efficiency is 1, and generalization is the 0 behind it. Without crossing the efficiency threshold, so-called generalization capability has no commercial significance." In his view, generalization only answers "can it be done?" but not "is it worth doing?" This year is called the first year of embodied AI commercial rollout, but most demonstrations have yet to reach the efficiency baseline required for commercialization. Geekplus's philosophy is clear: commercialization must pass the efficiency test.
This confidence stems from Geekplus's decade-plus of deep application in the field. Since its founding, Geekplus has positioned itself as an intelligent robotics company. "We just perfected the robot technology first, then targeted the warehousing and logistics industry, which has immense industrial value," Chen Xi says. Why warehousing? Because a warehouse holds millions of SKUs, covering most everyday items. The stable handling and grasping capabilities trained here become the "atomic actions" for foundational operations in future commercial or even home scenarios. All generalization capabilities for manipulating items can be trained in a warehouse.
The task that best forges these capabilities in a warehouse is the seemingly simple "picking." "Bin-picking" is called the "Holy Grail problem" in robotics. Conquering it lights up the entire capability tree. A single picking task tests four areas: vision must identify countless item shapes, physics must determine grasping angle and force, strategy must plan optimal paths and sequences, and fault tolerance must handle countless surprises like occlusion, deformation, and lighting changes. Industry consensus is that warehouse picking is the scarcest training ground for Physical AI, meeting all five conditions: high frequency, reality, clear feedback, transferability, and near-commercial closure. The prerequisite for satisfying these five conditions is having a business network operating in real-world tasks daily. Geekplus has exactly that. Operating in over 40 countries, with more than 1,700 projects deployed and serving over 950 international brands, its tens of millions of daily orders continuously feed its models, making them smarter over time.
Technical Routes Are Still in Selection, but the Flywheel Has Started Racing
What accelerates this flywheel is Geekplus's newly released Gravity 4D embodied model. "The biggest problem with traditional embodied models is their fragmentation. Models that understand language and semantics struggle to calculate physical movement, while models that predict physical movement lack high-level semantic understanding," points out Chen Chao, head of Geekplus's Embodied AI Large Model Team. Gravity uses a "dual-brain collaborative" architecture: the "cognitive brain" handles complex instructions and breaks down tasks, while the "action brain" performs physical "sandbox simulations" before acting. Unlike traditional visual large models that only predict "what the next frame looks like," Gravity 4D extracts 4D latent features to simultaneously learn future RGB appearance, 3D structure, and 3D motion. In Chen Chao's words, it "gives robots physical intuition," enabling them to understand the physical world like an engineer, not generate images like a director.
Regarding the industry's heated debate between VLA (Vision-Language-Action) and world models, Chen Xi remains remarkably calm: "The marathon of embodied AI is still in its early stages. Technical routes are not yet unified, and all are worth exploring. The actual capability gap between companies is currently small. The real dividing line isn't which technical route is chosen, but who enters real-world application scenarios first, runs a complete, deployable solution, and gets the data flywheel spinning." When the technical route converges, the marathon will enter its sprint phase. At that point, players with rapidly spinning flywheels and strong commercialization capabilities will hold the winning hand.
An Open Moat Grows Deeper
With the technological foundation in place, Geekplus cannot rely solely on itself to become an "embodied AI solution expert." At WAIC 2026, Geekplus launched the GINO ECO open ecosystem plan. It opens its highly reliable, low-cost hardware platform to embodied model companies and universities, collaborates with leading clients across industries to explore innovative scenarios, and partners with complementary hardware companies to create end-to-end, all-domain embodied workflows.
Someone might ask: sharing barriers—what happens to the moat? Chen Xi's answer: The embodied AI industry chain is very long, and one company cannot cover all aspects. Geekplus's role is as a bridge. Upstream model companies need real-world scenarios and data to validate their capabilities, and Geekplus can provide a stable hardware platform. Downstream end clients need complete solutions, and Geekplus can integrate partner products for delivery. "The industry has not yet reached competition; collaboration is more important than competition." The more open the ecosystem, the more embodied model companies, hardware firms, and scenario clients it attracts, making the Gravity framework increasingly irreplaceable as the technological foundation.
Embodied AI is a long-distance race, not a sprint. For 11 years, Geekplus has honed its fundamentals, relying not just on technology but on a complete commercial logic: anchoring on warehousing as the first scenario, defining commercialization thresholds with efficiency, and expanding its moat through an open ecosystem. As Chen Xi says, this marathon has only just begun, and Geekplus has been running on the path of solving problems.
Comments