Robot Implementation in ROS 2 Environment with Diverse Sensors, Interfaces, and SOM-based AI Processing

AMD Kria AI SOM as the "Brain," ADI Provides "Sensing and Neural Network" Architecture

Many robot development teams eventually face the same question.

"The semiconductor appears to have sufficient performance, the benchmark numbers are impressive, and the demo runs smoothly. So why does it still take 18 months to bring a deployable robot to market?"

The problem is rarely the processor alone. The issue lies in the elements surrounding the processor.

Today, the critical question for autonomous robot developers is not "which computing module delivers the highest processing performance (peak throughput)?" Rather, the core question is "how can we use promising semiconductor components to complete a functional, verified, and production-ready robot faster?"

■ From Microcontrollers to AI-Based Computing: A New Autonomous Robot Stack

Autonomous robot architecture is evolving. In the past, multiple microcontrollers (MCUs) were distributed across the system, each handling sensing, motion control, or communication functions. Now this approach is consolidating into a more centralized, AI-based robot computing architecture.

The AMD Kria™ AI Robotics Developer Platform reflects this transformation. This platform is a turnkey, open, fully integrated platform for autonomous robot development, supporting a wide range of robot applications from factory robots and autonomous mobile robots (AMRs) to mobile manipulators and humanoids. The Kria AI Robotics Developer Platform accelerates developers' transition from the concept stage to design and prototyping. Computing performance is provided by the Kria AI SOM (System-on-Module), which integrates CPU, integrated GPU (iGPU), NPU, and unified memory in a single device. Additionally, FPGA and ADI's sensing and connectivity technology stack are combined to complement the SOM's capabilities.

The platform is supported by the AMD Robotics Software Suite, based on AMD ROCm software. This suite provides an open-source runtime stack that includes familiar tools for developers: Linux, ROS 2, PyTorch, TensorFlow, Docker, and hardware-accelerated awareness libraries.

The familiarity of this development environment is critical. For many robot development projects, x86 Linux is not a compromise but the most suitable development environment. Development speed can be further accelerated when development teams do not need to invest time in learning a new proprietary SDK.

■ Integration Challenges That Delay Robot Development Projects

Even with a high-performance computing platform, significant time is spent on integration work not directly related to autonomous feature development. While this issue is well known, it must be solved from scratch with each new hardware design. Five common challenges encountered in development projects are as follows:

1) Wiring Complexity for Multiple Camera Connections

Robots that rely on visual perception often require four, six, or more camera inputs. Managing cables from so many sensors while maintaining signal integrity, temporal synchronization, and per-camera power delivery while conveying all signals to a central computing node becomes a complex mechanical and electrical design challenge in itself.

2) Localization in Real-World Environments

In environments where camera and LiDAR performance degrades—such as dimly lit warehouses, reflective floors, or featureless corridors—inertial sensing becomes essential. Integrating a high-reliability IMU with precise timestamping capabilities and an ROS 2 driver is not merely a component procurement matter. Calibration and software integration work are also necessary.

3) Depth Data Processing Burden

Depth sensors generate dense 3D data requiring real-time filtering, alignment, and fusion with other sensor data. Without a hardware-accelerated pipeline, these tasks compete directly with navigation and path planning tasks for host CPU computing resources.

4) Motion Control Where Safety Is Critical

Precision motion control cannot be achieved simply by transmitting speed commands over a bus. Motor control interfaces, isolated fieldbus connections, encoder feedback, and current sensing must be integrated into one system and validated under operational environmental conditions including vibration and temperature variations.

5) ROS 2 Readiness Time

One of the most overlooked aspects of robot development projects is the time required to make hardware subsystems functional in ROS 2. Every sensor and actuator requires drivers, hardware bindings, ROS 2 nodes, validated message types, and timing verification. Teams building systems based on generic boards must develop all of this themselves.

As a result, development teams invest considerable time in component selection, carrier board design, Linux driver porting, sensor timing verification, and ROS 2 integration for camera interfaces, repeating this process for each subsystem.

■ Providing an Immediately Applicable Nervous System Beyond a Simple Brain

Through the combination of the Kria AI Robotics Developer Platform and ADI solutions, AMD presents a new approach to autonomous robot development. This collaboration does not simply optimize specific layers while leaving remaining integration work to developers; rather, it provides what robots actually need. In other words, it provides not just computing performance but a validated system-level foundation that functions like a nervous system.

The Kria AI SOM is combined with an open robot carrier card designed based on ADI technology. This carrier card provides FPGA and interfaces commonly used in robotics. These include high-speed camera connections based on gigabit multimedia serial link (GMSL™) validated in the automotive industry, inertial sensing, CAN-FD, RS-485, gigabit and multi-gigabit Ethernet, Automotive Audio Bus (A2B) for low-latency audio transmission, battery management connectivity, and motor control extension capabilities.

Using a standardized COM-HPC architecture, developers do not need to redesign custom carrier boards for each project. If computing requirements change, only the SOM needs to be replaced while the rest of the platform remains intact.

As autonomous robots advance to the stage of deployment in actual operational environments, safety and reliability are no longer optional. Robots must operate even in environments where visual capabilities are degraded, and this is where ADI's ADIS16607 IMU shines. According to ADI's sensor fusion research, combining an IMU as an auxiliary sensor with vision systems can significantly enhance system robustness across major datasets.

Computing performance likewise cannot be evaluated solely by TOPS figures. Robot workloads are highly complex. They must support not only low-precision transformer inference for perception models but also classical high-precision algorithms, numerical optimization, calibration, and development workflows. The x86 GPU and NPU architecture integrated into the AMD Kria AI SOM is designed to handle both types of workloads.

ADI provides the physical interfaces necessary for robot implementation on this platform. These include highly reliable inertial sensing, consistent depth data processing, long-range camera links that integrate multiple video inputs through a single coaxial cable, isolated industrial communication for motor control, and reference designs for 48V robot power battery management.

■ Starting from a Validated Platform, Not from Scratch

The core value this platform provides is clear: developers do not need to start in a bare-bones Linux environment with no preparation.

The AMD Kria AI Robotics Developer Platform provides pre-validated hardware and software platform with integrated ADI drivers. Interface configuration is already defined, the ROS 2 environment is ready, and procedures for transitioning from development kits to production SOMs and actual robot systems are documented. Developers can focus on autonomous feature development from day one without needing to spend initial weeks on carrier board bring-up and environment setup. For many projects, this time savings can amount to several months.

■ Future Robot Development Requires Even More Complete Systems

What constrains the advancement of robots deployed in today's warehouses, factories, and logistics facilities is not computing performance itself. Rather, the time and cost required to build systems around that computing are the larger limiting factors. While there is always room for higher AI performance, the software stack supporting it has long been built by integrating multiple discrete technologies, making integration difficult.

ADI and AMD are building a platform to address these integration challenges. This platform is a nervous system for robots designed so that validated computing, sensing, connectivity, motor control, and ROS 2 software work together organically from the start. As development teams transition from prototype to production stages, ADI and AMD support accelerating that transition. For more information, visit the AMD Kria AI Robotics Developer Platform website (https://www.amd.com/en/products/system-on-modules/kria/ai/robotics-developer-platform.html).