Skip to main content

Introduction

AI Robot is a hardware-software co-designed solution for next-generation intelligent robots. Built on the self-developed K1 RISC-V processor and a proprietary operating system, it integrates on-device AI inference, large language model (LLM) support, and high-performance control capabilities. The solution is widely applicable to service robots, quadruped robots, drones, robotic arms, and other robotic platforms.

The core system, ROS2_LXQT, is based on the self-developed Bianbu OS as its operating system foundation, with ROS2 as the runtime core. It integrates the multimedia middleware MediaEngine, the RVV-optimized libraries (RVV-Opt Libs), and the Robot Development Kit (RDK) to form a complete robotic application infrastructure. The system comes with a rich set of built-in AI models and toolchains, supporting common robotic functions such as speech recognition, visual perception, task orchestration, and path planning, along with ready-to-use reference examples for rapid deployment.

Platform Vision​

Bianbu Robot aims to promote the adoption of RISC-V architecture in robotics, accelerate the development of a domestic intelligent robot ecosystem, empower intelligent perception and natural interaction capabilities, and help developers rapidly deploy next-generation smart terminals.

System Architecture​

The Bianbu Robot solution integrates large language model (LLM) capabilities for natural language interaction and intelligent multi-task orchestration. It supports a wide range of intelligent terminal applications, including drones, cleaning robots, quadrupeds, robotic arms, and AI-powered consumer robotics.

At the system level, the Bianbu Robot solution is based on the Bianbu ROS robot operating system and provides a complete reference platform for AI robotics. Bianbu ROS is built on the proprietary Bianbu OS, integrates deeply optimized robotic middleware and SDKs, and includes the multimedia middleware JDK, high-performance computing libraries (HPC Libs), and the BRDK development kit.

RDK (ROS Dev Kits)​

RDK provides deep abstraction and modular encapsulation based on ROS2, offering open-source reference solutions aligned with AI robotics product requirements.

  • Integrated AI development toolchain Includes built-in quantization toolchains, foundational pretrained model libraries (Model Zoo), and sample code. Supports mainstream model formats and deployment workflows, ready to use out of the box.
  • Adapted for mainstream open-source large models Natively supports DeepSeek, Qwen, the Hugging Face Smol* series, and more. Integrates the large model orchestration framework LocalAI with OpenAI API-compatible frontend interfaces for seamless integration with existing ecosystems.
  • Complete speech processing solution Provides product-grade ASR (speech recognition) and TTS (speech synthesis) model libraries, multi-microphone array preprocessing, sound source localization SDKs, noise suppression, and echo cancellation modules — covering the full speech interaction chain from sensing to communication.
  • Open-source ROS2 service components Offers a wide range of ROS2 packages, including vision perception, voice interaction, sensor integration, embodied motor control, localization and navigation, and LLM task orchestration, supporting modular integration and customized development.
  • Reference application pipeline examples Provides end-to-end references covering data acquisition, processing, modeling, and interaction to help developers quickly build complete robotic application systems.

MediaEngine​

MediaEngine provides highly abstracted data pipelines based on the underlying OS, enabling modular functionality and high code reuse so developers can focus on building high-level pipelines without worrying about low-level data scheduling and system details.

  • IO HUB data hub Compatible with various camera devices and communication protocols, supports zero-copy data transfer, minimizes system bandwidth consumption, and significantly reduces power overhead.
  • Hardware-accelerated pipelines Fully leverages on-chip hardware acceleration units in pipeline flows, supporting high-performance image operations such as scaling, rotation, cropping, and blending.
  • Modular programming model Supports block-style programming across the entire flow, including data acquisition, encoding, model inference, watermarking, video display, and multi-stream republishing, for flexible composition and rapid development.
  • Low-latency direct display mechanism Provides flexible direct video rendering strategies, optimizes framebuffer paths, significantly reduces video pipeline latency, and meets real-time application demands.
  • OpenCL-accelerated library support Integrates a wide range of image enhancement and stabilization algorithms, including distortion correction and electronic image stabilization (EIS), accelerated by the AI CPU for efficient image processing.

RVV-Opt Libs (RISC-V Vector Optimization Libraries)​

High-performance computing is critical in robotic perception and control tasks. Bianbu Robot provides deeply optimized compute engines to maximize the hardware potential of the K1 chip.

  • Basic computation libraries Integrates mainstream foundational libraries such as OpenCV for computer vision, OpenVML for math, and Eigen/OpenBLAS for linear algebra. These libraries are vectorized and deeply optimized for the RISC-V platform to fully utilize the 256-bit RISC-V Vector Extension (RVV).
  • AI operator libraries Built upon custom AI instructions following the IME (Intelligent Matrix Engine) standard to form meta-operator-level AI operator libraries. By calling standard operator APIs, developers can harness powerful AI performance.
  • AI-accelerated inference backend Deeply adapted for the RISC-V platform with operator fusion, multithreaded scheduling, memory optimization, and dynamic SARM partitioning. Seamlessly integrates with ONNXRuntime, Llama.cpp, and other mainstream inference frameworks.
  • High-performance computing language Based on a RISC-V-adapted version of OpenAI Triton, developers can write Python code to achieve near-peak hardware performance. It supports rapid migration of mainstream CUDA inference tasks to the RISC-V platform, significantly lowering adaptation costs and accelerating deployment.

Solution Highlights​

The Bianbu Robot solution focuses on speech algorithms, AI vision, and large model applications, offering out-of-the-box modular services tightly coupled with chip capabilities to help clients accelerate solution deployment. SDK source code is open to support developer innovation.

  • Supports a rich Model Zoo Covers large model applications, speech processing, and traditional visual AI models.
  • Large model applications Compatible with Qwen2 series, DeepSeek-1.5b, SomlLM2, SomlVLM, SomlVLA series. Supports large model conversational speech interaction with end-to-end latency under 1.5 seconds.
  • Speech processing Covers VAD, ASR, TTS, DOA, NS, and multi-microphone sound source localization.
  • Traditional AI vision Offers optimized models for object detection, classification, tracking, and more, supporting over 1000 application scenarios.
  • Localization and mapping Supports laser navigation, monocular VSLAM, and stereo visual odometry solutions.
  • Supports Linux PreemptRT Compatible with industrial control protocols such as EtherCAT, CAN, and RS485/232, meeting real-time control requirements.
  • Low-latency video pipeline Supports up to 1080P@120FPS encoding, making it ideal for drone video latency requirements.
  • Complete toolchain Includes optimized GCC/Clang compilers, quantization tools, and large model fine-tuning tools.

Solution Advantages​

Current robotic terminals heavily rely on connectivity, but privacy protection is a key challenge. Bianbu Robot effectively addresses the core challenges of deploying large models:

  • Development flexibility Linux SDK system source code is fully open, with rich debugging tools to support custom algorithm development and application extension.
  • Cost advantage The K1 performs baseline AI inference on the edge, reducing dependency on cloud servers and lowering total system operation costs.
  • Performance balance Combines lightweight edge inference with high-performance cloud computing — ensuring local smart model upgrades while avoiding latency from cloud-only solutions, enabling always-on smart responses.
  • Enhanced privacy Handles sensitive tasks (e.g., ASR) locally using on-device compute, complying with privacy laws across countries and solving regulatory concerns in large model applications.

Feedback​

For issues or suggestions, please visit the project community:

🔗 https://gitee.com/bianbu/brdk-doc/issues