DiGua Robot
Abstract: High-computing-power chips are only the first hurdle for the commercial deployment of humanoid robots on production lines.
In recent years, the "brainpower" of humanoid robots has been rapidly expanding. NVIDIA's Jetson AGX Orin, released in 2021, delivers a peak computing power of 275 TOPS; the Jetson Thor pushes low-precision sparse computing power to 2070 TFLOPS; in China, DiGua Robot's Sunrise S600 offers 560 TOPS@INT8; XPeng's IRON is equipped with three self-developed Turing chips, with a publicly disclosed total computing power reaching 2250 TOPS.
Judging solely by computing power, the processing capacity for humanoid robots seems somewhat over-provisioned already.
However, once it comes to real-world scenario deployment, merely comparing computing power is no longer sufficient.
On July 15, 2026, DiGua Robot disclosed at the Sunrise S600 media communication meeting that the Sunrise S600 has entered the high-performance robot computing market with 560 TOPS. To verify real-world workloads, DiGua Robot demonstrated algorithms such as spatial perception, MultiCAM SLAM, VO-DP+, HoloBrain, and WM-LOCO. This embodied AI computing platform has already collaborated with over 20 robot and embodied model companies, and completed collaborative adaptation with more than 100 software, hardware, and solution partners. This progress reflects that domestic high-computing-power platforms have entered the customer testing phase, but cooperation, chip bring-up, real-machine adaptation, mass production verification, and batch delivery are different stages, and the numbers cannot be directly equated.
Hu Chunxu, Vice President of Developer Ecosystem at D-Robotics
"In the past era of showing off skills, a single backflip or a successful grasp was enough to attract attention; but now we have entered the real era of productivity, which requires robots to operate stably and continuously in their environment, possess high reliability, withstand outdoor cold and indoor heat, and run stably under various working conditions," Hu Chunxu, Vice President of Developer Ecosystem at D-Robotics, told eefocus reporters, noting that the computing power competition for humanoid robots has entered its second stage. Being able to run VLA, VLM, or large language models only indicates that the computing platform has the potential to carry next-generation algorithms. Whether the model can run with low latency within the robot's power budget, whether perception, planning, and control can run concurrently over the long term, whether the robot can keep up with the production pace, and whether the deployment cost can achieve a positive ROI all need to be verified item by item at real workstations.
Humanoid Robot Computing Power Competition: Why Can't We Just Look at TOPS?
First, a concept must be clarified: the AI performance published by chip manufacturers is usually based on specific precision, sparsity rates, and peak conditions.
What truly affects the robot's experience is "effective computing power." How many frames the same model can run at the target precision, how long the end-to-end latency is, whether multiple cameras, LiDAR, and force feedback data compete for memory when running concurrently, whether simultaneous VLA inference and motion control will cause jitter, and how long the chip can maintain peak performance under the robot's thermal dissipation conditions—these metrics are much closer to actual usage scenarios than a simple comparison of TOPS.
Therefore, when the author asked whether the 560 TOPS computing power achieved by the Sunrise S600 is already sufficient for humanoid robots,
Wang Cong, CEO of D-Robotics
Regarding this question, Wang Cong, CEO of D-Robotics, frankly stated, "Currently, there is no standard in the industry for how large the chip computing power should be, nor has anyone told me how large my algorithm needs to run... So this question has a different answer every year."
In fact, because embodied AI algorithms are still changing rapidly, it is difficult to solidify the computing power requirements for chips in a short time. For example, this generation of robots may distribute visual perception, localization, and motion control across multiple computing boards, while the next generation of products aims to integrate world models, task planning, and control loops into a single platform. Model parameters continue to increase, while low-precision quantization reduces the inference burden; the number of sensors is rising, and algorithms are also reducing intermediate modules through end-to-end architectures.
What humanoid robots need is not an isolated "high-computing-power chip," but a computing system covering cloud training, edge inference, and real-time control.
The cloud is responsible for model training, data generation, and simulation; the edge high-computing-power platform handles visual perception, environmental understanding, task planning, and multimodal interaction; and the real-time control unit processes joint, motor, and actuator commands. Whether these three layers can collaborate smoothly directly affects the robot's response speed, motion stability, and task success rate. Simply increasing the TOPS of the robot body cannot automatically solve problems such as control latency, system power consumption, thermal dissipation, battery life, and mass production.
The above figure classifies platforms into "high-computing-power brain platforms" and "lightweight main control platforms." Although the peak computing power of different platforms varies, they cannot be simply ranked by computing power.
Because computational precision, sparsity rate, model structure, chip utilization rate, and test conditions differ, the actual inference performance corresponding to the same computing power figure may vary greatly. Even if two chips have similar TOPS, they may exhibit completely different latency, power consumption, and throughput performance when running visual models, language models, world models, or motion control algorithms.
For robot manufacturers, what is truly valuable is the effective computing power that can be continuously invoked. Humanoid robots need to process multiple streams of data from cameras, depth cameras, LiDAR, IMUs, and force sensors simultaneously, while also completing model inference, task scheduling, communication, and safety control. If the memory bandwidth is insufficient, the computing units may wait for data for a long time; if the software stack cannot fully invoke the GPU, NPU, and CPU, peak computing power is also difficult to translate into overall machine performance. In addition, the edge platform must meet the requirements of long-term operation, fast wake-up, low-latency control, and safe response in abnormal states. These requirements collectively determine whether a chip can truly enter robot products.
Looking specifically at each computing power chip manufacturer, NVIDIA's core advantage remains general-purpose computing and the software ecosystem. Orin has formed relatively mature commercialized modules, and developers can use CUDA, JetPack, TensorRT, and Isaac to complete model development, deployment, and simulation. Models trained on servers are also easier to migrate to the robot edge. Thor further enhances low-precision inference capabilities, memory capacity, and bandwidth, and integrates with robot development systems such as GR00T, making it suitable for teams that need to quickly verify multimodal models and end-to-end algorithms.
The cost of this solution is equally clear. Orin's publicly disclosed power consumption ranges from 15 to 60W, while Thor reaches 40 to 130W. Every additional dozen watts on the computing platform will be transmitted to battery capacity, thermal dissipation structure, overall machine weight, and battery life. The procurement cost of high-performance modules will also affect the overall material cost of the robot. In the R&D stage, general-purpose platforms help shorten the distance from algorithms to real machines; after entering mass production, OEMs need to re-evaluate performance redundancy, supply stability, and unit cost. NVIDIA can maintain the high-end general-purpose development baseline, but its share in the mass production stage will depend on robot companies' tolerance for power consumption and cost.
Compared to NVIDIA, domestic chipmaker D-Robotics is cutting in from the directions of local toolchains, software-hardware collaboration, and real-time control. The Sunrise S600 uses a 4-core Nash BPU for AI inference, an 18-core Cortex-A78AE for task scheduling and general-purpose computing, and a 6-core Cortex-R52+ for safe real-time control, with a memory bandwidth reaching 204.8GB/s. Its design philosophy is to divide the labor of model inference, system tasks, and control loops within the same SoC (System on Chip), reducing data exchange and system integration pressure between multiple computing modules.
At the software level, S600 covers development, training, quantization, deployment, and real-machine debugging through RDK Studio, RoboGo, and Moss Agent Engine. Compared to simply announcing TOPS, this platform-based approach is closer to the actual needs of robot companies. Chip manufacturers not only need to provide hardware but also help customers migrate models, adapt sensors, integrate operating systems, and solve problems in real-machine debugging. The next step for S600 is to prove whether these tools can be reused across different customers and robot models, gradually shifting project cooperation from joint development to standardized delivery.
Qualcomm's IQ-9075 and Black Sesame Technologies' SesameX Liora continue some capabilities of automotive and industrial computing platforms. The IQ-9075 integrates an AI accelerator, Kryo CPU, and real-time MCU (Microcontroller Unit) into a 4nm SoC, emphasizing industrial temperature, long-term supply, and real-time subsystems, possessing the foundation to extend from AMRs and industrial robots to humanoid robots. Liora targets world models, end-to-end control, and multimodal perception, attempting to bring automotive-grade heterogeneous computing experience into embodied AI. Both types of platforms have entry points in reliability, software-hardware integration, and control collaboration, but currently, publicly disclosed mass delivery data for humanoid robots remains limited, making them more suitable to be seen as challengers entering the validation stage.
XPeng represents the vertical integration route of OEMs developing their own chips. The self-developed Turing AI (Artificial Intelligence) chip can be jointly optimized around its own models, sensors, data, and robot body, and the software and hardware teams can also iterate according to the same product rhythm. Its risk lies in the fact that R&D investment, tape-out costs, toolchain maintenance, and chip economies of scale must all be borne by OEM sales. XPeng has announced that IRON uses three Turing AI chips and a mass production target by the end of 2026, but it ultimately needs to be verified by actual delivery. The specific model, computing power, and mass delivery data of Tesla Optimus's current edge chips have not been fully disclosed, so they should not be forcibly included in parameter comparisons.
Rockchip, Intel, and AMD are located at other positions in the robot computing power chain: RK3588, with its lower cost, mature development resources, and supply foundation, can undertake main control, visual processing, or a lightweight "cerebellum"; Intel Core Ultra relies on x86 compatibility to cut into industrial control, robot host computers, and general-purpose computing; AMD Ryzen AI, Kria, and other platforms support edge inference, prototype validation, and some low-latency tasks through CPU, GPU, NPU, or FPGA heterogeneous computing. These products are not in direct competition with the "brain" platforms of humanoid robots, but choose to occupy important positions in control, connection, vision, or auxiliary computing.
What Four Thresholds Must Be Crossed for the Commercial Mass Production of Humanoid Robots?
Current shipment status of mainstream humanoid robots, Source: Compiled by eefocus Research Institute
2026 is called the "first year of commercialization" by the humanoid robot industry, and the growth in shipments is just one manifestation. A more substantive change is that robots are beginning to step out of laboratories and exhibition halls, entering factories and specific business processes, and undergoing joint testing of reliability, cost, delivery capability, and return on investment. As the whole machine shifts from technical demonstration to product delivery, the computing power platform, as the base for the robot's "brain" and "cerebellum," has also entered the stage of commercial validation.
Hu Chunxu, Vice President of Developer Ecosystem at D-Robotics, believes that the mass production of embodied AI is jointly driven by four forces: capital is beginning to make long-term investments; technology bases such as algorithm platforms, sensors, and hardware are steadily maturing; the demand for improving operational efficiency in various scenarios is accelerating; and the collaborative capacity of the upstream and downstream of the industrial chain is steadily advancing. However, forming a consensus on mass production does not mean that mass production problems have been solved. There is still a huge engineering gap for robots to move from laboratory prototypes to stable operations in factories, and model deployment, effective computing power, hardware reliability, and production line returns are all indispensable.
From public cases, the commercial deployment of computing power platforms roughly goes through four stages. First, the chip or module completes commercialization, such as Jetson Orin and Thor; second, robots equipped with related platforms enter the developer market, such as Unitree G1 EDU offering Orin as an option; then, customers complete real-machine adaptation, scenario testing, or mass production validation, such as Sunrise S600 cooperating with multiple robot companies; and finally, clear batch delivery and large-scale application. Although XPeng IRON has proposed a mass production target by the end of 2026, it is currently still in the planning stage.
These four stages cannot be directly equated. The stable supply of chip modules does not mean that the robots equipped with them are already running in batches in factories; completing chip bring-up or real-machine adaptation only indicates that the technical route is basically feasible. To truly enter mass production, it must also undergo multiple tests such as orders, overall machine yield rate, supply chain organization, production line pace, and delivery costs.
Taking the Sunrise S600 launched by D-Robotics as an example, the currently disclosed partners include Tashi Zhihang, UBTECH, Beijing Humanoid Robot Innovation Center, Humanoid Robot (Shanghai) Co., Ltd., Qianxun Intelligence, Independent Variable Robot, Zhi Zai Wujie, and Pacini, among others. The progress of different projects is not consistent; some have completed computing power adaptation, some are still in real-machine testing, and some models have announced subsequent mass production validation or batch delivery plans. Therefore, a more accurate judgment at this stage is that S600 has entered the multi-customer adaptation and mass production validation stage, and more data is needed to prove large-scale commercial delivery.
Wang Cong, CEO of D-Robotics, remains cautious about this: "By this year, I feel we have entered the stage where 'some scenarios can be implemented, and true product-level applications are beginning to emerge.'" Currently, relatively easy-to-deploy tasks are mainly concentrated in scenarios with clear boundaries such as material transport, loading and unloading, sorting, and inspection. The industry's assessment criteria are shifting from whether the robot "can move" to "whether it can work stably." To achieve cross-factory and cross-industry replication, computing power platforms must also cross four thresholds.
The first threshold is model deployment and software migration. The industry is exploring deploying large models such as VLA and world models to the robot edge. If the models mainly run in the cloud, network latency and jitter will affect the real-time performance and reliability of control; if transferred to the edge, they must face constraints such as power consumption, cost, memory capacity, and bandwidth. Migrating models from training frameworks to heterogeneous platforms such as BPUs and NPUs also requires completing operator adaptation, quantization, precision regression, memory scheduling, and performance analysis.
Wang Cong stated, "Switching to different chip architectures inevitably requires investing more personnel in algorithm migration and adaptation, which is a normal phenomenon when changing technical routes." Chip performance can be rapidly improved in one or two generations of products, but operator libraries, development tools, debugging capabilities, technical documentation, and engineering talent require long-term accumulation. If domestic platforms rely on resident engineers for project-by-project adaptation for a long time, the more customers there are, the greater the delivery pressure may be. Therefore, the maturity of the toolchain must ultimately be measured by the migration cycle, manpower invested, and subsequent maintenance costs.
The second threshold is transforming peak computing power into sustainably invocable effective computing power. Humanoid robots need to process data from cameras, LiDAR, IMUs, encoders, and force sensors simultaneously, complete perception, localization, task planning, and action generation, and then send the results to the control loops of dozens of joints. Being able to run the VLA model only covers a part of the entire computing chain. When model parameters, visual features, and multi-stream sensor data are frequently transferred, memory bandwidth may become a bottleneck earlier than the AI (Artificial Intelligence) computing units.
Public data shows that the memory bandwidth of Orin 64GB and S600 are both 204.8GB/s, Thor is increased to 273GB/s, while Qualcomm and Black Sesame related platforms lack whole-machine memory bandwidth data under a unified caliber. Without a unified baseline model, computational precision, batch size, and power consumption conditions, TOPS can only indicate the theoretical upper limit. The inference throughput of specified models, first-frame latency, multi-task concurrency capability, and whether frequency scaling occurs after long-term operation are closer to the actual performance that robot manufacturers need to evaluate.
The third threshold is coordinating the different time scales of the "brain" and the "cerebellum." Large models can take tens of milliseconds or even longer to complete high-level decision-making, while joint and motor control requires shorter and more deterministic responses. S600 is configured with a Cortex-R52+ real-time control core, and IQ-9075 also sets up a real-time MCU (Microcontroller Unit) subsystem, reflecting that chip manufacturers are integrating perception decision-making and safe real-time control into the same SoC (System on Chip) or computing platform.
This integration helps reduce the number of boards and data round trips, lowering the complexity of the whole machine, but it also poses new reliability requirements. When visual reasoning, language interaction, path planning, and joint control run simultaneously, whether the platform can guarantee the determinism of real-time tasks needs to be verified through long-term, high-load operation of real machines. A smooth movement at an exhibition cannot replace reliability tests under abnormal states, temperature changes, and continuous loads. Mass production also requires the standardization of PCBs (Printed Circuit Boards) and interfaces, and the completion of certifications for high-temperature resistance, impact resistance, and high load. These tasks need to be jointly promoted by chip, body, sensor, controller, and manufacturing companies.
The fourth threshold is production line success rate, operational pace, and return on investment. Factories require robots to perform repetitive tasks for a long time, safely recover after recognition errors, grasp failures, or path blockages, and maintain a consistent pace with the preceding and subsequent processes. Adding tens of milliseconds to model inference, occasional recognition errors, or replanning paths can all translate into production efficiency losses.
Chen Tongqing, Co-founder of Tashi Zhihang, pointed out, "Therefore, POC validates technical feasibility, and production line validation validates commercial feasibility. Industrialization is not just scaling up the POC a bit, but re-answering the three accounts of yield rate, capacity, and ROI." After entering the factory, robots also need to connect to WMS, MES, and safety systems, adapting to different workstations, materials, lighting conditions, and personnel flow. If adding a single workstation requires recollecting a large amount of data and debugging for weeks, scale replication will be difficult to establish; if downtime recovery relies on manual labor for a long time, the saved labor costs may also be offset by O&M investments.
Therefore, the "first year of commercialization" represented by 2026 is closer to the starting point for humanoid robots to shift from prototype competition to product and production line validation. For domestic computing power platforms such as S600, obtaining adaptation opportunities from multiple robot companies is only the first step.
Chen Tongqing, Co-founder of Tashi Zhihang
From Computing Power to Ecosystem: Competition in a Higher Dimension
The next question in the computing power competition for humanoid robots is whether domestic platforms can replicate NVIDIA's CUDA ecosystem.
The barriers of CUDA go far beyond GPU performance, also including compilers, operator libraries, development frameworks, debugging tools, technical documentation, developer communities, and engineering talent. NVIDIA also connects cloud training, simulation validation, and edge deployment through JetPack, Isaac, and commercialized modules. Around this system, developers have accumulated a large amount of code, tools, and project experience. Switching chip architectures often requires re-handling operator support, model quantization, precision regression, memory scheduling, and performance optimization. Even if domestic chips have comparable peak computing power, it is difficult to replicate the software stickiness formed by CUDA over the long term in the short term.
However, the humanoid robot industry has not yet formed a unified hardware architecture and development standards, which leaves a market window for domestic SoCs (System on Chip). Robots need to run VLA, world models, and multimodal perception algorithms, as well as process multi-stream sensor data, system task scheduling, and joint real-time control. General-purpose GPUs have an advantage in the development ecosystem, but they may not be able to simultaneously meet the needs of power consumption, cost, control determinism, and localized services. Domestic platforms can first cut in from edge inference, real-time control, local supply, and scenario adaptation, establish advantages in specific robot tasks, and then gradually expand the developer base.
The opportunity for domestic manufacturers such as D-Robotics also lies here. It is introduced that Sunrise S600 integrates BPU, CPU, real-time control cores, and multi-sensor interfaces into the same SoC (System on Chip), and covers model quantization, deployment, performance debugging, and real-machine development through tools such as RDK Studio and RoboGo. Providing algorithms such as spatial perception, MultiCAM SLAM, VO-DP+, HoloBrain, and WM-LOCO, on the one hand, provides customers with reference capabilities, and on the other hand, is also used to verify whether multi-stream vision, Transformer operators, and motion control loops can run stably on the platform. What DiGua provides is no longer just a chip, but a robot computing base covering hardware, operating systems, development tools, and basic algorithms.
But open platforms still have to face migration costs. Wang Cong, CEO of D-Robotics, stated, "Switching to different chip architectures inevitably requires investing more personnel in algorithm migration and adaptation, which is a normal phenomenon when changing technical routes." Teams with edge quantization and autonomous driving development experience can usually complete the migration faster; teams that mainly engaged in algorithm prototype development in the past and lack quantization deployment experience will face higher onboarding costs. To achieve higher efficiency on NVIDIA platforms, deep quantization and performance optimization are also required, it's just that a large number of developers are already familiar with its workflow, so the related costs no longer seem prominent.
This also determines the core problem D-Robotics needs to solve: how to precipitate one-time engineering support into standardized tools. Hu Chunxu introduced that DiGua's software system basically adopts an open approach, equipped with technical documentation and Agent development tools, hoping to lower the threshold for secondary development. He cited an example where the team recently helped a customer adapt a previously unsupported camera sensor. The traditional method required an experienced ISP engineer to spend 5 to 10 days on tuning. After using the Agent tool, it took about two hours to obtain an image effect with color differences imperceptible to the human eye. This case shows the potential of automated tools to improve adaptation efficiency, but its universality still needs to be verified in more sensor, algorithm, and robot projects.
On this basis, DiGua has proposed the positioning of a "mother ecosystem." According to Hu Chunxu's explanation, the "mother ecosystem" does not mean that DiGua completes all software and hardware links, but provides an industry-universal standardized platform, allowing body, sensor, actuator, controller, operating system, and application development companies to collaboratively develop on the same base. DiGua hopes to become the "greatest common divisor" in the field of robotics through chips, modules, operating systems, development tools, basic algorithms, and standard interfaces.
It is revealed that over the past years, D-Robotics has connected more than 100,000 developers globally, participated in building millions of robot products, landed over a hundred models, and the platform has accumulated more than 400 intelligent algorithms and applications; computing power products cover 5 TOPS to 560 TOPS, and provide different forms such as chips, modules, and development kits. These accumulations mainly come from previous consumer-grade, service-oriented, and other robot products. Whether they can be smoothly migrated to embodied AI and humanoid robots depends on the degree of reuse of different technical assets.
Existing cases provide preliminary validation. Vita Dynamics' Vbot super robot dog uses Sunrise S100 to run natural language interaction and environmental perception models on the edge; Accelerating Evolution's small and medium-sized humanoid robots use Sunrise S100P to undertake brain computing and process binocular perception through RDK X5. Both parties also plan to explore edge applications of multimodal robots based on S600. The above projects show that DiGua's platform has entered specific products, but moving from a single project to cross-customer replication still requires a longer period of mass production validation.
Currently, D-Robotics has completed deep adaptation based on Sunrise S600 with hardware partners such as Bosch Semiconductor and GigaDevice, completed solution integration with perception manufacturers such as SmartSens and OV, and cooperated with automotive industry Tier 1 robot teams such as Luxshare Precision, ZhiXing Technology, Jinmai Electronics, and Huaqin Technology to launch mass-production-level robot controller solutions based on Sunrise S600. Through advance collaborative adaptation upstream and downstream, it effectively reduces customers' whole-machine integration, debugging, and mass production trial-and-error costs.
Finally, the author believes that in the next two to three years, as VLA, world models, and multimodal perception enter the robot body, manufacturer competition will extend from computing power to memory capacity, bandwidth, heterogeneous computing, real-time control, and development efficiency. Rather than building a CUDA ecosystem in the field of humanoid robots, a more realistic direction is to first become a universal base in domestic robot computing platforms, establishing differentiated advantages in edge inference, real-time control, software-hardware collaboration, and local engineering services. From computing power to ecosystem, domestic humanoid robot computing chips will usher in competition in a higher dimension.