Since 2025, the BOM (Bill of Materials) cost of mainstream humanoid robot manufacturers has dropped to under 200,000 RMB per unit, with some manufacturers claiming to have entered the stage of small-batch delivery at the thousand-unit level. But another set of data is equally noteworthy: according to industry research, among the humanoid robots currently in hand, more than 70% are still stuck at the "skill demonstration" stage of showroom demos, going up and down stairs, and carrying mineral water bottles, while the proportion that truly enter factory production lines for continuous operation is less than 5%. The hardware cost curve has already been pushed down; what is holding back scale-up is insteadsoftware stack—this is almost the most underestimated link in this round of theembodied intelligencewave.
Over the past five years, the hardware iteration speed of humanoid robots has been astonishingly fast. Core components such as harmonic reducers, planetary roller screws, and high-torque-density joint motors have completed the leap in the domestic supply chain from "relying on imports" to "self-sufficiency," directly driving the overall cost of humanoid robots from the million level down to the 200,000 level. The robots of manufacturers such as Unitree, AgiBot, and Galbot are no longer weak in motion capability either. Movements such as backflips, running, and walking on complex terrain, which were still laboratory demonstrations two years ago, are now standard at product launches.
But the problem lies precisely in the huge gap between "can walk" and "can work." The human definition of "working" is continuously performing tasks for hours with clear quality requirements and tolerance for sudden anomalies—for example, plugging and unplugging wiring harnesses in a 3C production line, or sorting packages of different sizes in a warehouse. Such tasks require not some stunning single-point action, but a stable combination of an entire set of software capabilities: environmental perception, task planning, force control precision, exception recovery, and human-robot collaboration safety. These capabilities have not yet formed a mature supply chain like "joint motors," and manufacturers are all building their own software stacks while fighting their own battles.
If you take apart a humanoid robot and look at its software system, you will find that it is actually an extremely complex real-time computing platform. From the bottom up, at least three layers are current common technical bottlenecks in the industry.
The first layer is real-time control and communication.When humanoid robots perform fine operations such as plugging and unplugging or tightening screws, the response frequency requirement for the force control loop is usually above 1 kHz. This means that the force sensor signal of a set of joints must complete acquisition, filtering, closed-loop calculation, and command issuance within milliseconds. Traditional industrial robots solve this with real-time industrial computers plus dedicated buses. Because humanoid robots have more degrees of freedom and looser structures, they often have to assemble fieldbuses such as EtherCAT and CAN on general-purpose computing platforms, and the engineering tuning difficulty for real-time performance and determinism far exceeds imagination. This is where many companies doing robot software are most likely to stumble.
The second layer is embodied large models and the data closed loop.The "brain" of humanoid robots now generally adopts a layered architecture of "VLA (vision-language-action) large models" + high-frequency traditional controllers. The large model is responsible for slow high-level decision-making, and the traditional controller is responsible for fast low-level execution. This architecture is very smooth in demos, but once it enters a real production line, data becomes the biggest bottleneck: high-quality robot operation data is far scarcer than text data. Collecting a set of operation data for "plugging and unplugging wiring harnesses" requires motion capture equipment, teleoperation platforms, and long periods of manual annotation, with costs often reaching tens of thousands of RMB. Without data, the generalization ability of the VLA model cannot improve; if it cannot generalize, it cannot adapt to the ever-changing working conditions on the production line.
The third layer is deployment, operations and maintenance, and OTA iteration.This is a link that is easily overlooked yet extremely fatal. Industrial scenarios have extremely low tolerance for downtime, while the software version iteration frequency of humanoid robots is far higher than that of traditional equipment—large model weight updates, control parameter tuning, and perception algorithm upgrades mean new versions are almost pushed every month. How to complete firmware upgrades without interrupting production line operations, how to ensure that old tasks do not regress after upgrades, and how to perform remote monitoring and status diagnosis on hundreds of devices—these operations and maintenance capabilities directly determine whether a robot project can recover its costs. In a sense, the final implementation of embodied intelligence no longer competes on algorithms, but on engineering and operations and maintenance systems.
The supply chain data is very telling. According to public information, by 2024 the domestic localization rate of core components for humanoid robots had already exceeded 60%, and by the end of 2025 this figure was still continuing to rise; at the same time, the inflection point has appeared where leading manufacturers' shipment targets have shifted from "hundred-unit demonstrations" to "thousand-unit commercialization." But rising shipment volumes will precisely amplify software problems—demonstration machines can tolerate crashes and restarts, but mass-produced machines on production lines cannot.
A trend even more worthy of attention is that embodied intelligence is shifting from a "movement-centered display logic" to an "operation-first value logic." Whoever first creates a robot that can stably complete some specific industrial operation and has a data flywheel effect will be the first to establish barriers in vertical scenarios. This will bring a new round of division of labor: manufacturers good at hardware provide the body, teams good at software provide a general operation platform, and what is truly scarce is the system integration capability to twist the two together on the production line.
For China'ssoftware developmentindustry, the opportunities brought by humanoid robots go far beyond the "robot body" itself. Low-level real-time control software, the embodied data platform in the middle, and industrial applications and operations and maintenance systems at the top—each link requires a large amount of professional software engineering investment. In a city like Shenzhen where manufacturing and the software industry are highly concentrated,IoTcombined with robot software is becoming a new growth point, from smart factories tosmart community solutionsthe common needs of device access, data management, and remote operations and maintenance are being accelerated and amplified by the robot wave.
Technology itself is not the goal. Using technology to bring robots from showrooms into production lines and create real business value is the ultimate decisive factor in this track. ForShenzhen software developmentteams, the earlier they enter the engineering link of embodied intelligence, the more proactively they can occupy a position in the next round of industrial division of labor.
Room 1409, Building C, Shenzhen Century Industrial Center, No. 743 Zhoushi Road, Hezhou Community, Hangcheng Sub-district, Bao'an District, Shenzhen
