Digua Robotics releases its first single-SoC computing and control integrated robot development kit and has reached cooperation with over 20 leading customers | The Latest Frontline

Author | Huang Nan

Editor | Yuan Silai

On June 11th, Sweet Potato Robotics released the industry's first single-SoC computing and control integrated robot development kit, RDK S100. With a humanoid big and small brain architecture design, it can support the efficient collaboration of large and small embodied AI models, connect the action closed-loop of "perception - decision - control", and drive the collaborative evolution of the big and small brains of embodied robots.

The RDK S100 provides a variety of peripheral interfaces and full - link development infrastructure support for software - hardware collaboration and end - cloud integration. It can help embodied robot developers quickly build products and deploy them in multiple scenarios, accelerating the large - scale implementation of embodied AI.

RDK S100

Currently, there are two major directions in the evolution of embodied AI technology: end - to - end and hierarchical decision - making model architectures. The latter has strong interpretability and relatively low implementation difficulty. Based on these advantages, the big and small brain architecture for hierarchical decision - making models is a more feasible solution for robots to achieve embodied AI. It requires a single - SoC platform that can balance low power consumption and diverse computing power combinations to provide computing support.

The RDK S100 integrates a CPU + BPU + MCU on a single SoC. While reducing hardware costs, it can improve overall performance, enabling the "big brain" that is best at decision - making and the "small brain" that is best at control to achieve super - heterogeneous and efficient collaboration, meeting the computing requirements of both perception reasoning and real - time motion control, and realizing the trinity of "perception + decision + action".

A quadruped robot dog equipped with RDK S100

Based on the humanoid big and small brain collaborative heterogeneous architecture design, the RDK S100 can support the dynamic fusion and seamless switching of large and small models, connecting the complete closed - loop from perception to execution to handle various complex and changeable real - world scenario tasks. Among them, large models can provide accurate data support for robot decision - making and operation by deploying popular models such as visual/point - cloud detection, LLM, and VLM. Small models provide stable and reliable high - speed and high - precision motion control by deploying various motion control models. Coupled with the close cooperation between the MCU and BPU, it can reduce the CPU load by 80% and further improve the robot's reaction speed.

Taking fetching a cup as an example, when a human gives an instruction, the CPU and BPU visual semantic detection models of the RDK S100 will start to operate. They will efficiently process the received instruction and the visual information collected by the robot in real - time synchronously. Combined with the current state of the robot, the BPU will perform motion control reasoning to complete the decision - making and planning of the task. Then, the MCU will cooperate with the BPU in real - time based on the planning results for motion control processing, output action instructions, and precisely guide the robot to perform the action of fetching the cup.

A dual - arm robot equipped with RDK S100

Officially revealed, the RDK S100 provides a variety of peripheral interfaces, including 40PIN, MCU, M.2, USB, Flash - connect Type - C, camera expansion, JTAG, etc., to meet various needs such as research, testing, and prototype building. At the same time, Sweet Potato Robotics also provides full - link development infrastructure for software - hardware collaboration and end - cloud integration for the RDK S100, including multiple AI model data closed - loop solutions, hardware acceleration function packages, model adaptation and tuning solutions, and more than 200 open - source algorithms and application examples. This helps developers and customers complete product building faster, achieve multi - scenario adaptation of the RDK S100, and accelerate product implementation.

Hu Chunxu, the vice - president of the developer ecosystem at Sweet Potato Robotics, pointed out in an interview with 36Kr and other media that according to different industry dimensions, current robots can be roughly divided into traditional robots, including industrial robots and sweeping robots. The industry concentration is relatively high. Although the market scale is still growing, the competitive landscape is basically set.

The second type is new - form robots, which are mainly developed for specific scenario requirements, including pool cleaning robots, lawn mowing robots, ball - picking robots, and home companion robots. These products provide innovative solutions to user pain points in niche scenarios through intelligent technology, representing a new direction for robot applications.

The third type is embodied robots, typically represented by humanoid robots and large quadruped robots. These products have general - purpose functions and need to handle complex environmental perception and decision - making tasks, so their computing power requirements are significantly higher than the previous two types.

There are significant differences in computing power requirements among different types of robots. Hu Chunxu said that in addition to the RDK S100, Sweet Potato Robotics will also develop more boards with different computing power specifications to adapt to diverse application scenarios.

Hu Chunxu, Vice - President of the Developer Ecosystem at Sweet Potato Robotics

Currently, Sweet Potato Robotics has established partnerships with more than 20 leading embodied AI customers. At the same time, the RDK S100 is being tested by more than 50 customers. It is priced at 2,799 yuan in the market and can be applied to multi - legged robots, small bipedal robots, semi - humanoid robots, quadruped robots, LeRobot robotic arms, humanoid robots, as well as BEV detection, multi - channel video detection, and other products.

In the first quarter of 2026, Sweet Potato Robotics will launch more high - computing - power products to support more high - level embodied robot application scenarios.