PUBLISHER: ResearchInChina | PRODUCT CODE: 2168688
PUBLISHER: ResearchInChina | PRODUCT CODE: 2168688
Research on Robot Controllers: In-depth Review of Architecture Roadmaps, Chip Selection and Supplier Landscape
Robot Controllers (Cerebrum & Cerebellum) Research Report, 2026 analyzes and summarizes the following six dimensions: 1. definition, functions, architecture and classification of robot controllers (cerebrum & cerebellum); 2. development history, integration solutions and challenges of cerebrum-cerebellum convergence for robot controllers; 3. development trends of robot controllers; 4. chip applications for robot controllers; 5. controller deployment solutions adopted by domestic and overseas robot OEMs; ⑥ robot controller suppliers and their product portfolios.
Robot controllers are undergoing a four-stage architectural evolution, transitioning from discrete heterogeneous / on-board heterogeneous architectures to integrated computing-control architectures.
The controller hardware architecture evolves through decoupling followed by re-integration. In the centralized era before 2000, a single DSP or MCU handled trajectory interpolation and point logic, with no independent AI computing power. Starting from 2015, the industry entered the discrete heterogeneous phase: the AI computing board (cerebrum) and the independent motion control board (cerebellum) communicate via Gigabit Ethernet. The cerebrum and cerebellum are physically separated, at the cost of cross-board latency and jitter. Since 2025, the industry has moved into the integrated computing-control phase. The NPU and high-performance cores run Linux, while real-time cores run RTOS. On-chip shared memory replaces external buses. The native convergence architecture remains in the prototype stage, targeting a perception-decision-execution closed-loop latency below 10 milliseconds.
It is worth noting that discrete heterogeneous architectures are not obsolete. In the short term, discrete heterogeneous solutions will still be adopted for research prototypes, industrial and high-safety scenarios, as they enable complete fault domain isolation. During the pre-mass-production R&D phase, most manufacturers adopt a compromise "PCB on-board heterogeneous" design, integrating the AI SoC cerebrum and real-time MCU cerebellum on the same mainboard. Over the next three to five years, discrete heterogeneous, on-board heterogeneous and integrated computing-control architectures will coexist rather than replace one another in a single transition.
Chip Selection by Domestic Robot OEMs: Three Approaches for Three Types of Robot Platforms
Consumer-grade quadruped robots adopt domestic SoCs for cost reduction (e.g., Rockchip RK3588), with extreme cost sensitivity as the priority. Industrial-grade quadruped robots deploy modular computing hardware (e.g., Jetson Orin NX, D-Robotics RDK S100), prioritizing SLAM and environmental perception. Humanoid robots adopt a separated cerebrum-cerebellum design paired with high-end chips (e.g., AGX Orin with 275 TOPS, flagship Jetson Thor with 2070 FP4 TFLOPS) to deliver redundant computing power for end-to-end VLA. In-house designed chips have also started to be deployed on robots.
Computing power tiers for commercialization are divided into three brackets:
Entry-tier for domestic cost reduction, 0.5-6 TOPS, represented by RK3588. Large models run on the cloud, while the edge only handles reflex control, with high-volume shipment as the target.
Mid-tier for mainstream mass production, 70-130 TOPS, represented by Orin NX and RDK S100. Task planning and perception run locally on the edge, targeting large-scale delivery.
High-tier flagship , 275-2070 TOPS, represented by AGX Orin, Jetson Thor and in-house chips. Local execution of VLA foundation models and end-to-end pipelines with reserved computing redundancy.
Three Tracks for Overseas Robot OEMs
The overseas market is split into three distinct tracks. Consumer/service humanoid robots (1X NEO, Figure 03) focus on low power consumption, quiet operation and cloud collaboration, mostly relying on off-the-shelf third-party platforms. Industrial/logistics humanoid robots (Agility Digit, Apptronik Apollo, Boston Dynamics Atlas, Oversonic RoBee) prioritize real-time performance and force control for payloads, with layered decoupling being the most mature architecture. The automotive/autonomous driving migration faction (Tesla Optimus, Vin Motion Motion2) ports the ADAS computing stack directly into robots, featuring the highest level of vertical integration.
When selecting chip architectures, overseas OEMs face a three-way trade-off essentially between "ecosystem dependency vs vertical integration": out-source Jetson platforms (mature CUDA ecosystem, out-of-box deployment, but constrained by supply and cost, prone to homogenization); vertically integrated self-developed SoCs (controllable cost, optimized hardware-software co-design and data closed-loop, yet requiring massive R&D investment and forming a closed ecosystem); heterogeneous off-the-shelf combinations (flexible selection, low power consumption and guaranteed real-time performance, yet heavy integration workload).
Three Major Camps of Robot Controller Suppliers, with Industrial Automation as the Core Base
The positioning of controller suppliers is largely shaped by their original business background, falling into three main categories.
From the perspective of commercial maturity, the dividing line among suppliers has become clear.
The first tier has achieved large-scale shipments: SEER Robotics delivered more than 8,000 units in H1 2026 (up 80% year-on-year), with new orders worth CNY 467 million (+60%) and customers across 43 overseas countries and regions. Ruixinxing's XMC 300 has served over 80 manufacturers, with annual shipment growth exceeding 200%.
The second tier is in mass production delivery: iMotion Automotive Technology's iRC 100 has secured orders from leading robot companies; Luxshare Precision CORTEX 100 and Huaqin's self-developed main controllers are being rolled out sequentially.
Key Conclusions
Development Trend of Robot Controllers (Cerebrum & Cerebellum)
Policies, Regulations and Industry Standards for Robot Controllers
Compute Chip Landscape for Robot Cerebrum & Cerebellum
Controller Selection Characteristics of Domestic Robot OEMs
Architecture Roadmap and Chip Supply of Domestic Robot OEMs
Three Commercial Tiers of Controller Compute for Domestic Robot OEMs
Product Iteration and Risks of Controllers from Domestic Robot OEMs
Summary of Controller Selection by Overseas Robot OEMs
Chip Roadmap and Selection Trends of Overseas Robot OEMs
Three Paths for Controller Suppliers to Enter Embodied Intelligence
Supplier Chip Solutions and Commercial Maturity
Next 3-5 Years: From "Purchased & Assembled" to "Integrated Hardware-Software"