Must the "big brain" and "little brain" of a humanoid robot be separated?

Edited by Taylor From Gasgoo
Main productsNew Energy Vehicles Module Products: DC Fast Charging Modules, Energy Storage and Charging Modules, On-board Power Modules Monitoring Products: Charging Control Modules, Charging Cloud Platform System Products: Standard DC EV Chargers, Split-type Charging Systems, Fully Liquid-cooled Supercharging Systems, AC EV Chargers, V2G EV Chargers
Main productsPrecision Castings, Component Blanks, Machined Parts

Gasgoo Munich-A humanoid robot reaches for a cup on the table. To succeed, it must simultaneously accomplish two tasks.

The "big brain" determines the cup's location and the grip angle. The "little brain," meanwhile, constantly refines the movement based on joint status, center-of-gravity shifts, and tactile feedback. If the task or environment changes, the big brain can re-plan the approach, while motion-level deviations are corrected in real-time by the little brain and underlying closed loops.

Historically, these two categories of tasks were handled by separate computing platforms. The big brain typically relied on high-performance SoCs or compute modules for environmental perception, task comprehension, and motion planning. The little brain used real-time processors for motion control—usually calculating and updating target positions, velocities, or torques for whole-body joints on a millisecond-scale control cycle—before passing instructions to the joint actuators.

Now, an increasing number of chipmakers are attempting to integrate robot perception, decision-making, and control into a single SoC, pushing for "calculation-control integration" of the big and little brains.

Does this signal that the dual-chip approach is being displaced by single-SoC solutions?

One Chip or Two: Manufacturers Are Betting on Both

CASBOT recently unveiled its latest collaboration with AMD.

According to CASBOT, its CASBOT 02 robot demonstrated low-latency, on-device multimodal interaction at the 2026 World Robot Congress. It achieved this using an external AMD Ryzen AI embedded X100 series compute unit paired with a localized large model inference engine.

image.png

Image Source: CASBOT

Leveraging the X100's on-device compute power, the CASBOT 02 runs local large models and Agent scheduling directly on the robot body. This supports multi-lingual, continuous multi-turn dialogue and allows the robot to execute motion skills based on conversation semantics.

Visual, auditory, and proprioceptive data also contribute to task understanding and judgment, reducing the robot's reliance on cloud communication.

However, for its next generation, CASBOT has no intention of keeping all tasks on a single chip. Instead, it plans to adopt a dual-processor configuration using AMD's Ryzen AI embedded X100 and P100 series, assigning the "big brain" and "little brain" roles separately.

Specifically, CASBOT's next humanoid robot plans to upgrade the Ryzen AI X100 from an external unit to an internal "brain" in the chest cavity. It will handle multimodal perception, complex model inference, task understanding, Agent scheduling, and high-level decision-making.

The P100 series will ensure motion control and communication real-time performance, while reserving AI expansion capabilities for next-generation perception and adaptive control algorithms. This supports a base version of the robot capable of adaptive motion control relying solely on the "little brain."

At a time when single-SoC solutions are a hot topic, CASBOT's decision to go the other way with a dual-processor setup is telling.

In fact, CASBOT is not alone in adopting a dual-processor architecture for the big and little brains in the humanoid robot field.

According to a review of publicly disclosed schemes by Gasgoo Embodied Intelligence, numerous teams over the past two years have adopted physically separated or logically layered designs for their humanoid robots' big and little brains—though the division of labor varies.

A typical example is Figure 02, which houses two GPUs running Helix's System 2 (S2) and System 1 (S1) respectively. S2 runs visual-language models at 7–9 Hz for scene understanding and high-level planning, while S1 outputs continuous upper-body motion control at 200 Hz.

Black Sesame Intelligence previously partnered with Academician Liu Sheng's team at Wuhan University to build a big-little brain platform for the "Tianwen" humanoid robot. The big brain platform selected the Huashan A2000 to provide compute power for multimodal information processing and complex decision models. The little brain platform is based on the Wudang C1236, using a time-sharing operating system and hardware isolation technology to allow AI computing and control tasks to run in parallel.

Yet, over the past year, publicly disclosed single-SoC calculation-control integration schemes have also continued to increase, moving from development platforms to full machine adaptation and product launches.

image.png

Image Source: Vbot

For instance, Vbot's Vbot ATOM has confirmed it will carry Dagu Robot's Xuri S100P single-SoC calculation-control integration architecture, achieving a "perception-decision-action" trinity. According to the plan, ATOM is expected to begin deliveries by the end of 2026 and is currently open for reservation.

Beyond Vbot ATOM, Astribot's cable-driven AI robot, the T1, has also completed Xuri S100 adaptation and key link verification. Both parties plan to continue joint optimization and push for a full rollout of the solution within the year.

ACE Robotics's self-developed wheel-arm lifting fulfillment robot, meanwhile, selected Huixi Technology's big-little brain fusion chip, the Huixi Technology Embodied R1 MAX. The robot has already entered trial operation at a Kunlun Haoke convenience store (PetroChina Zhenxing Store).

Even UBTech has entered the chip-making fray. Its first project targets a "big-little brain integrated" SoC for robots, with a development goal to "completely surpass NVIDIA Thor."

Currently, UBTech's joint venture with Maxsilicon and Fenglong—Xixuan Chuangzhi—is advancing the architecture design of its self-developed main control chip. The company plans to provide a development and verification platform to industry partners in 2027 for adaptation and testing, with the official launch of the first SoC product and accompanying embodied intelligence solutions slated for 2028.

Overall, chipmakers are primarily driving the platform definition behind these recent single-SoC disclosures, while OEMs participate through functional partitioning and body requirements to jointly push solutions toward machine adaptation.

This shares similarities with the centralization of automotive E/E architectures in recent years: both aim to reduce computing units and shorten data paths.

The difference lies in the fact that robots must handle dynamic balance and whole-body real-time control, meaning they cannot simply replicate the evolution of automotive domain controllers.

What Problems Can Two Chips Solve?

Consolidating the big and little brains into a single SoC is a clear trend, so why are some manufacturers still choosing dual processors?

"CASBOT’s 'big brain + little brain' architecture stems from its own needs for modularity and functional partitioning," explains Feng Yi, Vice President of AMD's Greater China Embedded Division.

In his view, the key to robot development isn't pursuing the extreme in a single metric, but balancing performance, power consumption, real-time response, and reliability.

Architecturally, the value of dual processors lies not just in added compute, but in providing clear physical isolation for time, resources, and safety boundaries.

First is the time boundary. The big and little brains face different computational loads, operating cycles, and latency determinism requirements. A dual-chip architecture can better meet these needs.

The big brain runs large models like visual-language models (VLMs) or VLA. How long a single inference takes often depends on input images, context length, and current load—variable speeds are the norm.

The little brain manages body balance, trajectory tracking, and whole-body coordination. It must operate at a strict, fixed frequency. Every cycle, it must read sensors, calculate targets, and issue commands on schedule. It requires deterministic latency—it cannot rely on merely being fast on average.

For example, when a robot moves objects, a slight delay in the big brain re-planning the task might just mean a slower decision. But the little brain cannot stop for a moment; joint status, center-of-gravity shifts, and foot forces must all be processed in time according to their respective control cycles.

Two chips provide relatively independent computing resources and scheduling environments. This helps ensure the big brain runs at the pace of inference, while the little brain runs at the pace of the control cycle.

image.png

Image Source: Astribot

Second is the resource boundary. Physical separation reduces the direct encroachment of the big brain on the little brain's computing resources.

AI inference doesn't just occupy the GPU; image processing, model scheduling, and sensor data movement also consume CPU, cache, and memory bandwidth. Even if the priority of motion control is raised, as long as the hardware is shared, resources like memory bandwidth and on-chip buses can still be starved, affecting real-time performance.

After physical separation, the little brain gains relatively independent CPU, memory, and communication interfaces. Adding a camera or swapping in a larger model on the big brain side won't directly squeeze out motion control resources.

In short, a separated architecture turns resource contention between two systems into collaboration across a clear interface.

Third is the safety boundary. If the big brain fails, the little brain can serve as a safety net.

A dual-chip architecture can place verification and limiting functions on the little brain side, checking speed, torque, and pose constraints on targets issued by the big brain. With independent state perception and degraded control capabilities, the little brain can intercept dangerous movements—maintaining posture, decelerating, or stopping safely—even if the big brain outputs anomalies or stops responding.

Of course, two chips don't automatically guarantee safety. Without designs for heartbeat detection, command timeouts, safety monitoring, and emergency stop circuits, physical separation alone won't prevent dangerous actions.

From this perspective, a dual-chip scheme simply carves out an independent deployment space for safety mechanisms; true safety still relies on a complete control strategy.

Finally is the iteration boundary. Upgrading the big brain doesn't require touching the little brain.

Typically, perception models, VLA, and task planning algorithms update more frequently. Control firmware, however, requires repeated calibration around the specific body—involving joint characteristics, mass distribution, communication cycles, and stability verification. The iteration rhythms of the two are inconsistent.

If the two are tightly coupled, updating a model might require re-testing the entire control pipeline.

Once split, both sides collaborate via a relatively stable interface. Downstream, the big brain sends task targets, action primitives, or trajectories, which the little brain translates into whole-body coordinated movements before passing them to lower-level joint MCUs for closed-loop control. Upstream, the little brain can pre-process high-frequency body data like joint states and IMU readings, then upload only status summaries, anomaly events, or richer state features as needed by the upper-layer models. This reduces unnecessary data transmission.

Of course, specific command formats vary by manufacturer, but the logic of division is roughly the same: as long as the interface and motion constraints remain unchanged, the upper layer can swap models or adjust planning with minimal impact on the underlying motion control.

Single-SoC Integration Is a System-Level Calculation

There is no denying that a dual-chip architecture provides more favorable conditions for real-time control, resource isolation, and independent iteration of the big and little brains through physical boundaries.

But two computing platforms also bring additional costs for data transfer, hardware configuration, and development maintenance.

For products still in the prototype or small-batch delivery stage, the cost of two computing platforms may be acceptable. But once mass production begins, material costs, space, power consumption, and delivery cadence must all be factored into the overall machine cost. OEMs then need to reassess which independent boundaries are essential and which redundant configurations can be merged. The evolution of automotive E/E architectures from distributed to domain-centralized has already demonstrated this.

The appeal of a single SoC stems precisely from this recalculation of system costs.

优必选造芯片,把目标对准了英伟达Thor

Image Source: UBTech

The first consideration is shortening the data path.

Passing target commands like "move forward" or "grab cup" between two chips isn't hard. The challenge lies in the high-frequency closed loop from perception to control. After a visual model identifies an object's position, the little brain must continuously correct the movement based on joint status and force feedback. Transferring this data across devices usually involves data encapsulation, interface transmission, parsing, and time synchronization between two systems. The overhead depends on the communication mechanism and interface design.

By comparison, with unified memory, zero-copy, and reasonable scheduling, a single SoC can reduce cross-device data movement and lower communication latency and jitter.

Moreover, this shorter data path facilitates the exchange of execution deviations and contact status between perception models, motion policies, and traditional controllers.

However, it is worth noting that while single-SoC integration can improve data exchange, it does not mean large models can take over high-frequency joint servo loops. The lowest-level position, velocity, and current loops still require real-time control modules or joint drivers.

The second consideration is reducing hardware redundancy.

A "brain computer + little brain controller" setup involves more than just the main chips; it often includes two sets of memory, storage, power management, connectors, and cooling structures. For humanoid robots with limited torso space and battery capacity, this supporting hardware consumes significant weight and volume.

A single SoC can reduce some redundant configurations by integrating different computing units onto one platform. For example, Dagu Robot's Xuri S100P integrates CPU, BPU, and MCU on a single chip to achieve a "perception-decision-action" trinity. With 128 TOPS of on-device compute, it can handle concurrent VLA and VLM inference and motion control without relying on the cloud for real-time decision closed loops.

ScreenShot_2026-11-16_171941_084.png

Image Source: Intel

The 3rd Gen Intel® Core™ Ultra processor, leveraging the heterogeneous compute of CPU, GPU, and NPU—combined with Intel® TCC, real-time operating systems, CPU core binding, and cache resource isolation technologies—can also carry AI inference and deterministic real-time control on a single SoC. When compute power meets task demands, no additional independent AI accelerator card is needed. When cross-device real-time communication is involved, it can work with TSN to achieve precise time synchronization, keeping network latency and jitter within predictable ranges.

Upon entering mass production, reducing the main control board or compute module translates to "lightening the load" for procurement, assembly, and testing.

The third consideration is lowering development and maintenance complexity.

Separated architectures often involve multiple development environments simultaneously—x86 and Arm, Linux and real-time operating systems, GPU and MCU. If the model team changes an output format, the control team may have to adjust the communication protocol, increasing integration and verification workload.

Unifying the hardware platform is expected to reduce adaptation and integration work across different hardware platforms.

Of course, these benefits come with a premise: the single SoC must still retain necessary resource partitioning, real-time guarantees, and safety isolation.

Moreover, in terms of cost, a single chip is not inherently cheaper or more power-efficient. High-end heterogeneous SoCs are often expensive themselves, and running AI inference and motion control simultaneously concentrates thermal pressure. The hardware benefits of integration are reflected more in reducing redundant configurations than in a direct price drop for the main chip. Whether it offers better cost-performance requires a comprehensive comparison of data paths, thermal management, power, development verification, and long-term maintenance.

In this sense, the number of chips is not a scorecard for architectural superiority.

More than chip count, the industry needs to answer: Can motion control complete on time when models are under high load? Can the underlying layer limit dangerous actions if the AI system fails? While shortening data paths and reducing hardware redundancy, have new resource contention or single points of failure been introduced?

Conclusion

Just as the centralization of automotive E/E architectures didn't strip all real-time control of its independent boundaries, the same likely holds true for humanoid robots: computing platforms can centralize, but safety and real-time tasks still require deterministic guarantees.

However, robots must handle more frequent dynamic balancing, contact, and whole-body coordination. Their real-time closed loops and fault handling present constraints different from those of automobiles. Real-time kernels within a single SoC, or independent controllers in dual-chip architectures, are both seeking the right landing point for this boundary.

Ultimately, the industry's real competition isn't over chip count or integration level, but over who can enable a robot to think clearly and stand steady—at a lower system cost.

IncreaseSupplierCompany Insights
Location: China/Guangdong/Shenzhen/Nanshan District
Main productsNew Energy Vehicles Module Products: DC Fast Charging Modules, Energy Storage and Charging Modules, On-board Power Modules Monitoring Products: Charging Control Modules, Charging Cloud Platform System Products: Standard DC EV Chargers, Split-type Charging Systems, Fully Liquid-cooled Supercharging Systems, AC EV Chargers, V2G EV Chargers
ChipsSupplier
Location: China/Taiwan/Taiwan/New Taipei City
Main productsPrecision Castings, Component Blanks, Machined Parts

Gasgoo not only offers timely news and profound insight about China auto industry, but also help with business connection and expansion for suppliers and purchasers via multiple channels and methods. Buyer service: buyer-support@gasgoo.com Seller Service: seller-support@gasgoo.com

All Rights Reserved. Do not reproduce, copy and use the editorial content without permission. Contact us: autonews@gasgoo.com

Related Documents(3)

Rankings of smart cockpit component suppliers in China (Jan.-Sept. 2026).7z
China Passenger Vehicle Export Rankings for Jan.-Sept. 2026.7z
Rankings of ADAS component suppliers in China (Jan. - Sept. 2026).7z