Beijing releases a three-year action plan for embodied intelligent robots, explicitly naming support for RISC-V system-level intelligent computing chips
The Beijing Municipal Bureau of Economy and Information Technology issued the "Beijing Embodied Intelligent Robot Industry High-Quality Development Action Plan (2026–2028)," proposing to cultivate no fewer than 5 world-class and 10 domestically leading humanoid robot complete-machine enterprises by 2028, with the Beijing-Tianjin-Hebei region accounting for 25% of national complete-machine production capacity. Under its technological innovation source project, the plan supports tackling key chips such as RISC-V-based system-level intelligent computing chips, and requires complete-machine units to adapt and apply the relevant chips; it also proposes newly building embodied intelligence dedicated computing power clusters with a total scale exceeding 50,000 PFlops.
Why read this
AI score 58
The plan writes RISC-V into the main text of the special robot policy and ties it to complete-machine adaptation, offering a view of the device-side chip roadmap and implementation constraints.
Topics
Body · AI translation
Beijing Releases Three-Year Action Plan for Embodied Intelligence Industry, Explicitly Backing Domestic RISC-V Chips Source: Semiconductor Horizons Published: 2026-11-12 10:50 RISC-V Embodied Intelligence Generated Poster This is also a rare case in Beijing's robotics-specific policies of writing an instruction set architecture into the main text.
Recently, the Beijing Municipal Bureau of Economy and Information Technology issued the "Beijing Action Plan for High-Quality Development of the Embodied Intelligent Robot Industry (2026–2028)," effective from the date of issuance and valid until December 31, 2028. The plan proposes that by 2028, it will cultivate no fewer than 5 world-class and 10 domestically leading humanoid robot complete-machine enterprises, with embodied intelligence models, robot bodies, and key technologies all reaching world-class levels, the Beijing-Tianjin-Hebei region accounting for 25% of national complete-machine production capacity, and the supporting rate for key components exceeding 60%. Centered on this goal, the plan deploys six major projects: embodied intelligence leapfrogging, Beijing-Tianjin-Hebei industrial chain coordination, technological innovation sourcing, large-scale application promotion, industrial ecosystem leadership, and security governance taking the lead.
Among the six major projects, chips do not form a separate section but are embedded in the list of key core technology breakthroughs under the technological innovation sourcing project.
Article 9 of the plan states that support will be given to tackling core components such as high-end flexible tactile sensors, micro-precision reducers, and micro coreless motors, whose supply faces risks of being constrained by others, as well as key chips such as RISC-V-based system-level intelligent computing chips and high-dynamic-load N-channel MOSFETs, promoting the layout of all key chips and core components that should be laid out. The sentence immediately following is that complete-machine units will be supported in adapting and applying relevant components and chips, promoting the adaptation of all related products that can be adapted.
Read together, the policy's intent is clear. The first half addresses the supply side, requiring chips to be produced; the second half addresses the demand side, requiring complete-machine manufacturers to install the chips that have been produced. Chips are not an isolated technology breakthrough item but are bound into a closed loop of complete-machine adaptation. In Article 8 of the same project, on frontier technology layout, ultra-low-power intelligent computing chips also appear, pointing to scenarios with limited end-side power consumption.
Among the plan's quantitative targets, one set of numbers is easily overlooked: forming a complete-machine production capacity of 200,000 units per year in Beijing. Compared with the complete-machine performance indicators, the value of these 200,000 units becomes clearer. The plan requires the complete machine's rated load to reach more than 40% of its own weight, mean time between failures to exceed 2,000 hours, the annual failure rate of core motion components and end effectors to be below 10%, the designed service life of the complete machine to exceed 5 years, battery endurance of more than 4 hours under continuous operation conditions, and the creation of 3 to 5 complete-machine platforms whose products can be defined by software.
These indicators jointly point to one constraint: end-side computing power must complete multi-channel vision, force, and tactile fusion within a limited power budget, run inference of vision-language-action models in real time, and at the same time leave sufficient margin for control loops. The higher the load-to-weight ratio, the greater the joint current, the tighter the margin for power supply and heat dissipation, and the smaller the space left for the computing unit.
The training-side account has already been calculated in the plan. The document proposes newly building embodied intelligence dedicated computing clusters with a total scale exceeding 50,000 PFlops, and collecting more than 1 million hours of high-quality data in no fewer than 100 real scenarios. But training clusters cannot solve the chip inside the robot's body. 50,000 PFlops is a number in the data center; the end side depends on how many tokens can be run per watt and whether inference latency can be pressed down to the millisecond level.
This is also a rare case in Beijing's robotics-specific policies of writing an instruction set architecture into the main text. Choosing RISC-V as the architecture is not only about supply chain security. The open nature of RISC-V allows manufacturers to make custom extensions on top of the base instruction set, hardening vector operations, matrix multiply-add, and dedicated operators into the pipeline, which is a real source of efficiency for end-side inference of VLA models. By contrast, the licensing terms of closed architectures restrict such deep customization, and both cost and cycle are not under one's own control. In addition, RISC-V has no licensing fees or royalties; for robot chips with expected annual shipments in the hundreds of thousands and extreme sensitivity to BOM, this cost difference will be amplified.
China already has accumulated experience on this route. Manufacturers such as XuanTie, Nuclei, StarFive, ESWIN Computing, and SpacemiT have been laying out RISC-V processor cores and high-performance computing chips for years, and Beijing also has a corresponding enterprise base locally. By setting the support direction at the architecture level, the plan effectively gives these existing investments a clear landing scenario.
The plan uses "system-level intelligent computing chip," not processor core, nor AI accelerator. The threshold implied by these five characters is higher than many imagine. A system-level chip for humanoid robots usually needs to integrate an application processor, neural network acceleration unit, image signal processing, real-time control core on the same die, and also handle multiple high-speed interfaces and a functional safety island. The engineering workload on the software side is equally staggering: compilers, operator libraries, quantization tools, model deployment frameworks, mixed scheduling of real-time operating systems and general-purpose operating systems—if any link is missing, the chip cannot run mainstream models. At present, mainstream VLA models are almost all trained and validated within the CUDA ecosystem, and the conversion cost of migrating to a new architecture is often higher than the price difference of the hardware itself.
Functional safety is another hurdle. Robots need to work in human-machine collaborative environments, and the plan specifically sets up a security governance taking-the-lead project, proposing hardware overload protection, active emergency braking, decision safety audit traceability, and full-process management with one code per machine. These requirements will ultimately translate into chip-level failure detection, redundancy design, and certification costs.
This article is reproduced from media reports or online platforms and represents the author's personal stance or views. Our republication is solely for sharing and does not mean we approve of or endorse it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service promptly, and as a neutral platform service provider we will promptly correct, delete, or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations
Yunnan Germanium's controlled subsidiary participates in Lumentum Core Supplier Conference to discuss supply chain collaboration 2026-11-12 12:11
South Korea's chip exports in early November increased by 331.2% year on year 2026-11-12 10:37
Robin Li: The four-layer architecture of "chip, cloud, model, and agent" was gradually formed over ten years 2026-11-12 10:36