Chipmakers Want One Chip for a Robot’s Brain and Cerebellum. Makers Disagree

A humanoid reaching for a cup has to do two things at once. The brain decides where the cup is and how to grasp it. The cerebellum reads joint state, centre of gravity and hand force, then corrects the motion continuously. When the task or environment changes, the brain replans while the cerebellum fixes deviations in real time.

Those jobs used to run on separate platforms. A high-performance system-on-chip handled perception, task understanding and action planning, while a real-time processor updated joint position, speed and torque on a millisecond cycle. Chipmakers now talk about merging both into one SoC, a design the industry calls compute-control integration.

A humanoid robot striding forward against a plain orange studio background
Robot makers are split on whether perception and real-time control belong on one chip (Source: Gasgoo)

The case for two processors

Zhongke Huiling disclosed its latest work with AMD. At the 2026 World Robot Conference, its CASBOT 02 used an external AMD Ryzen AI Embedded X100 computing unit with a local model inference engine to deliver low-latency multimodal interaction on the machine itself. The robot ran a local model and agent scheduling, held multilingual multi-turn conversations and called robot skills from the meaning of a dialogue. Vision, hearing and body state all fed task understanding, reducing reliance on the cloud.

For the next generation, the company is not putting everything on one chip. It plans an X100 series processor built into the chest as the brain, handling multimodal perception, complex inference, task understanding, agent scheduling and high-level decisions. A P100 series chip becomes the cerebellum, guaranteeing real-time motion control and communication while leaving room for adaptive control algorithms. The design also supports a basic version that runs adaptive motion control on the cerebellum alone.

CASBOT is not alone. Figure 02 carries two GPUs running Helix System 2 at 7 to 9 hertz for scene understanding and high-level planning, and System 1 at 200 hertz for continuous upper-body control. Black Sesame, working with an academic team led by Liu Sheng at Wuhan University, paired a Huashan A2000 brain for multimodal processing and complex decisions with a Wudang C1236 cerebellum using time-sharing and hardware isolation.

The single-chip answer

Disclosures of integrated designs have also grown. Weita Dynamics confirmed its Vbot ATOM will use D-Robotics’ Xuri S100P single-SoC architecture, combining perception, decision and action. Delivery is planned for the end of 2026 and pre-orders are open. Astribot has completed S100 adaptation and key-link validation on its rope-driven T1, with joint tuning under way. Daxiao Robotics chose Huixi Intelligent’s Embodied R1 MAX fusion chip for a wheeled arm robot now in trial at a convenience store.

UBTech is building its own chip, with an explicit goal of surpassing Nvidia’s Thor. The joint venture Xixuan Chuangzhi, formed with Muxi and Fenglong, is working on the architecture, plans a development platform for partners in 2027 and a first SoC in 2028.

A single chip cuts cost, power and board complexity. Two chips preserve hard separation between planning and millisecond control, which matters when a robot is carrying a load or working near people. The split is beginning to resemble an earlier argument in automotive electronics, where domain consolidation won in the end but only after safety-critical functions learned how to coexist.

Editor’s note: This is an adapted translation of the original Gasgoo report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.gasgoo.com/apps/e060bad9eca91db7f01403fc7120041f/robot/news/70478256-%E4%BA%BA%E5%BD%A2%E6%9C%BA%E5%99%A8%E4%BA%BA%E5%A4%A7%E5%B0%8F%E8%84%91-%E4%B8%80%E5%AE%9A%E8%A6%81-%E5%88%86%E5%AE%B6-%E5%90%97/.

Leave a comment