The internal structure of the system includes:

• 6x Arm Cortex-A78AE 2.0 GHz - application and system processing
• 1x BPU Nash - 128 TOPS (INT8) for CNN and Transformer model inference
• 4x Arm Cortex-R52+ 1.2 GHz - real-time motion control

The BPU is responsible for AI model inference, including motion control models, the MCU subsystem performs command interpolation and high-frequency communication with actuators, while the CPU manages the operating system and user applications. The architecture enables dynamic switching between perception and execution tasks as well as cooperation between large and small AI models within a single platform.

The integration of a 128 TOPS BPU with Cortex-R52+ real-time cores enables both perception and execution layers to operate within a single platform, without the need for external motion controllers.

The platform is suitable for applications such as:

• humanoid and semi-humanoid robots requiring synchronized perception and balance control
• quadruped and biped robots with dynamic motion stabilization
• manipulators and robotic arms performing precise grasping operations
• multi-channel video detection and image processing
• systems autonomous platforms operating in industrial or urban environments

The platform provides a complete development environment covering both hardware and software layers, enabling a rapid transition from prototype to production deployment. The architecture has been designed to support the full lifecycle of AI applications – from data acquisition, through model training, to optimization and deployment on an edge device.

The environment includes:

• hardware acceleration of AI models using the BPU Nash
• official ROS 2 package with hardware acceleration
• tools for model adaptation and tuning for specific application scenarios
• edge-cloud integration enabling implementation of a closed-loop data workflow
• support for popular AI models and ONNX operators

The applied solutions enable efficient utilization of the platform’s computing resources and reduce CPU overhead.

The platform enables parallel cooperation of the perception, decision-making, and execution layers within a single computing system. Data from visual sensors and user commands are processed by visual and semantic models (LLM/VLM), with inference carried out using CPU and BPU resources. The interpreted environment data and operator intent are then forwarded to the motion planning module, where the BPU performs computations related to trajectory generation and high-level motion control. The execution layer based on Cortex-R52+ MCU cores performs closed-loop control, handling command interpolation, actuator control, and high-frequency communication with the robot’s execution elements. This task division enables efficient distribution of computational loads, reduction of latency, and ensures predictable system response time, which is critical in robotic applications operating in dynamic environments.