After several years of visible hardware progress, Chinese robotics companies redirected much of their resources in early 2026 toward software that enables humanoid robots to operate usefully in real-world tasks. Investor interest in embodied AI — systems that learn from and act within the physical world — rose sharply this year, opening a new software-focused front in the US–China technology competition.
Funding and transactions
According to market researcher ITjuzi, Chinese companies working on robotics software that can interpret physical environments and learn autonomously completed 322 transactions in the first half of 2026 and raised 93.5 billion yuan (about $13 billion). That total is roughly five times the amount raised in the comparable period the previous year. As hardware prices are driven down by competition, manufacturers increasingly see the controlling software as the main source of future profit.
Players and strategies
Capital is flowing toward software developers responsible for robot control. Chinese startups such as X Square, Galaxea and Spirit AI have reached valuations around $3 billion. US startups — including Physical Intelligence, Skild AI and Generalist — retain advantages in advanced research and access to high-performance chips. For example, the Generalist GEN-1.5 model can perform about 60% of simple tasks after a 3–12 second demonstration.
Chinese firms, by contrast, leverage a strong domestic manufacturing base and close cooperation with industrial partners. X Square has demonstrated a method that allows robots to learn to handle objects from a single human video, without retraining the model for each task.
Data collection approaches and real-world testing
Unline text-trained chatbots, robots require examples from the physical world — from grasping objects to handling flexible cables — to learn. Chinese companies employ varied approaches to collect such data: Spirit AI hired 1,000 people to record everyday and industrial activities; ACE Robotics uses wearable sensors to measure movements and grip force of retail and hotel workers; and Robbyant, the robotics unit of Ant Group, trained a model on 160,000 hours of data, including 50,000 hours of human-operated robot motion, then tested it across 20 different robots from 17 brands.
Some firms, like DexForce, rely on simulated environments to prepare machines for precision industrial tasks. They report that certain tasks can be taught in six to eight hours with over 99.9% accuracy in simulation, though experts caution simulations still cannot perfectly reproduce real-world physical interactions.
Practical experiments already show promise. In a JoyIn Technology test, a robot that initially could not reach a high cabinet watched a human video, pushed a box under the cabinet and attempted to climb on it. In stricter industrial deployments, performance is higher: Spirit AI robots at battery maker CATL’s Chengzhou plant connect flexible wires with over 99% success in tests, and Robbyant machines assemble pharmacy orders in two Shanghai drugstores under staff supervision.
Headwinds: chips, capital markets and regulation
Growth in the Chinese sector faces headwinds. US export controls have reduced access to advanced Nvidia chips, and public fundraising is constrained by high valuations and market saturation that invite regulatory scrutiny. Developers are responding by improving model efficiency, using synthetic data, and turning to domestic chip suppliers such as Huawei to replace lost compute capacity.
Expectations and timeline for breakout
Many are looking to next-generation foundation models for a breakthrough. Unitree founder Vang Hszinghszing said in September that they obtained encouraging results testing OpenAI’s GPT-6 on robots, and he suggested GPT-7 could deliver a major advance if machines can complete 80% of tasks in unknown environments from spoken instructions. Robotics consultant Georg Stieler cautions that even if demos reach such levels, several years — not months — will likely be required before reliable, widespread commercial deployment.
Conclusion
In early 2026 the Chinese robotics ecosystem shifted noticeably toward embodied AI software, attracting substantial capital and producing promising real-world tests. Nonetheless, chip shortages, capital-market limits and regulatory pressures — together with the remaining technical challenges of transferring simulated or observed behavior to robust real-world operation — mean mainstream commercialization still faces significant hurdles.
