Do you also have this question: Why can industrial robotic arms perform tasks quickly and accurately, while robots that make pancakes or brew coffee are not only slow but also more prone to errors? Why do robots, which claim to possess embodied intelligence and can dynamically respond to the external world through perception, appear so "clumsy"?
If analyzed from a technical perspective, the reasons for the slow movements and sluggish responses of robots can be simply summarized in one word: high latency. The shorter the latency, the higher the chances of commercializing embodied intelligent robots.
Slow Reactions: Who Is Holding Them Back?
A qualified embodied intelligent robot is equipped with a complete closed-loop system comprising perception, decision-making, and execution. At the perception layer, the robot needs to perceive the world through visual sensors such as cameras and LiDAR, tactile sensors like electronic skin, and auditory sensors such as microphones. At the decision-making layer, the robot processes the acquired information with the help of edge processors and pre-trained world models. At the execution layer, the robot acts on the real world and changes the environmental state through motors, joints, robotic arms, and dexterous hands.
The reason why many embodied intelligent robots appear "clumsy" lies precisely in the varying degrees of processing latency present in the aforementioned stages.
In the pre-perception processing stage, operations involve the access and fusion of multi-sensor data, timestamp alignment of video streams, multi-source data synchronization, compression encoding, and spatial cropping. This stage inherently introduces a baseline fixed latency.
In the cognitive decision-making stage, the robot performs scene understanding, task reasoning, and behavior planning based on the Vision-Language-Action (VLA) large model. The latency in this stage mainly stems from the peak computing power of edge inference hardware and the computational complexity of the model structure, typically at the scale of hundreds of milliseconds, making it the primary component of system latency.
The robot's understanding of the world relies on the perfection of computational models. At this stage, there is still a significant gap between the capabilities of the World Model and human cognitive levels. Cui Zhuo, Vice President of ZTE Corporation and General Manager of the Humanoid Robot Product Line, told reporters from China Electronics News that the accuracy of open-source world models is only 60% to 70%. When the robot's understanding of the world is insufficient, ensuring the success of a single task requires invoking more local computing power, thereby consuming more computation time. Furthermore, for specific tasks such as recognizing transparent packaging bags or irregularly shaped workpieces, the algorithmic computational load is much heavier, making robots more prone to missed grasps, incorrect grasps, and even the production of large batches of defective products.
In the execution control stage, the robot sends the high-level action commands output by the VLA to the lower-level motion control system (such as the cerebellum controller), which completes the inverse kinematics solving and dynamic constraint calculation, thereby generating drive and control signals for the joints and dexterous hands. The latency in this stage depends on the bandwidth of the servo actuators, mechanical transmission characteristics, and the real-time control cycle.
The data transmission architecture of robots is also not streamlined enough.
At this stage, the three major technical stages of robots mentioned above are controlled by different processors mounted on the robot—the perception and acquisition module, the global decision-making brain, and the motion control module are hardware-independent, much like three people in different rooms needing to pass notes back and forth to communicate. The preprocessing of single-sensor data alone takes tens to hundreds of milliseconds, and handing over the tasks of recognition, decision-making, and planning to the VLA large model consumes another hundreds of milliseconds. Transporting multimodal data back and forth between different chips not only accumulates latency but also increases the overall power consumption of the machine and shortens the service life of servos and dexterous hands.
Xu Haijiang, Co-founder and Senior Vice President of Lingjing Zhiyuan, stated, "In current industry solutions, the perception-to-decision-making pipeline is fragmented, with all data relying on on-board processors for inference and computation. Completing tasks requires repeatedly transporting data from memory, which inevitably increases latency and power consumption."
Another important reason for the long processing time of robots is the lack of a unified latency testing benchmark among different components. Sensors, servo motors, computing chips, and the whole machine are usually independently developed by different enterprises. Each piece of hardware and software generates a tiny latency of several milliseconds. The industry has long lacked a unified latency testing benchmark, and the timing parameters of upstream and downstream components are mutually incompatible, which may also significantly increase the joint debugging and optimization costs for whole machine enterprises.
Controlling Latency: By What Standards?
If latency cannot be reduced, they will be eliminated by the market. Through interviews, reporters found that latency has become a rigid benchmark for whether humanoid robots will be eliminated by the market. "Demonstration prototypes" that can only perform demos in fixed, static environments and cannot adapt to dynamic real-world scenarios are being accelerated out of the market. In an interview with reporters from China Electronics News, Xu Haijiang stated that the two indicators of real-time response speed and task completion rate in complex dynamic scenarios are shifting from "bonus points" to the "passing line" in robot development.
In February this year, the "Standard System for Humanoid Robots and Embodied Intelligence (2026 Edition)" (hereinafter referred to as the "Standard"), jointly compiled by more than 120 research institutes, enterprises, and industry user units in the field organized by the Standardization Technical Committee for Humanoid Robots and Embodied Intelligence of the Ministry of Industry and Information Technology, was officially released. The "Standard" explicitly lists the dynamic task completion rate and end-to-end closed-loop response time as core indicators for equipment acceptance, and specifies that the end-to-end response time of the whole machine with a dual-cooperative architecture of the brain and cerebellum should be ≤200ms. The "Standard" is China's first top-level standard design covering the entire industrial chain and full life cycle of humanoid robots and embodied intelligence, and it also provides clear and rigid indicators for latency control in embodied intelligent robots.
In addition to end-to-end response time, Cui Zhuo also proposed a new perspective—there is an inverse relationship between action planning frequency and latency: the higher the planning frequency, the shorter the available time for a single planning cycle, and the lower the latency. A higher value for the action planning frequency indicates a faster robot response speed, making it more adaptable to scenarios with stricter latency requirements.
Cui Zhuo stated that the mainstream scenarios applicable to embodied intelligent robots can be roughly divided into three categories: the action planning frequency for commercial service robots is generally 10~20Hz; for precision industrial scenarios such as 3C electronics and automobile final assembly, the action planning frequency needs to reach 50Hz, with only a few milliseconds of processing window reserved for the hardware and software at each layer of perception, computation, transmission, and execution; although the overall latency threshold for home robots is relatively lower, when manipulating flexible and fragile items like clothing, eggs, and porcelain bowls, millisecond-level real-time force adjustment is still required. Cui Zhuo added that if wireless communication is adopted, the uplink transmission time limit is about 4 milliseconds and the downlink transmission time limit is 1 millisecond, leaving extremely little room for error.
High Latency: How to Reduce It?
During interviews, reporters learned that various enterprises are adopting different approaches to solve the latency problem in the task processing of embodied intelligent robots.
Cui Zhuo emphasized the path of algorithm optimization: the robot uses the world model to predict occlusion and object displacement scenarios in advance, and then uses the attention model to automatically filter out irrelevant targets such as passing pedestrians and idle clutter, enabling high-precision computing power only for objects at risk, which can significantly reduce the time consumed by repeated inference.
Lingjing Zhiyuan is targeting the robot architecture layer. The Dvorak super-heterogeneous computing architecture developed by Lingjing Zhiyuan adopts the design concept of "physical integration and logical separation." "In simple terms, it integrates the three major modules—the perception cortex, the decision-making brain, and the motion cerebellum—into the same chip hardware platform, eliminating the need to transmit large amounts of data across devices and fundamentally reducing transmission latency," Xu Haijiang introduced to reporters.
If the central master brain is analogous to the human prefrontal cortex that handles long-term thinking and logical planning, the terminal cooperative brain built into the dexterous hands and joints corresponds to the cerebellum, which governs human instinctive responses. By equipping lightweight edge VLA models, world models, and small brain-inspired emergency algorithms, humanoid robots can instantly adjust their movements locally when encountering sudden situations such as grasping slips or falling objects, without waiting for commands from the central large model.
This architecture exhibits similarities with the biological brain structure in terms of functional hierarchy. The system embeds high-level perception, cognition, and reasoning pathways capable of handling complex semantic and planning tasks, while also retaining low-level fast reflex pathways that can respond instantly to sudden stimuli. This mechanism is similar to human somatic reflexes: when our hand is burned by a cigarette butt, we instinctively pull away without waiting for the brain to perform conscious cognitive processing. Such an architecture can effectively enhance the real-time performance and robustness of the system in dynamic or sudden scenarios.
Ma Daolin, Founder and Chief Scientist of Qianjue Robotics, also affirmed the importance of the robot's local feedback closed loop. "The underlying control of a robot can receive control commands at thousands of hertz, but the inference frequency on the control model side is relatively low. There is still a large frequency difference between the two, which can easily cause latency at the decision-making layer and ultimately lead to task failure," Ma Daolin said. "By establishing a tactile-based local feedback closed loop at the end of the dexterous hand, grasping actions can be corrected more quickly without having to completely wait for the robot's brain to complete a round of global feedback."
Ma Daolin also proposed a technical solution to compensate for visual limitations with tactile sensing. "When the distance between the dexterous hand and the target object is narrowed to within 1 centimeter, relying solely on vision to judge the distance between the dexterous hand and the object is no longer accurate enough," Ma Daolin stated. "Fingertip tactile sensors with high sampling frequencies can complete the fine detection in the last 1 centimeter or several millimeters interval, identifying grip force, object deformation, and surface sliding trends; when encountering light occlusion or material deformation, touch can independently complete local recognition, saving the time consumed by repeated visual scanning and recognition."
Authors: Ji Xiaoting, He Ziyi (Intern) | Editor: Qiu Jiangyong | Art Editor: Ma Liya | Supervisor: Lian Xiaodong