.jpg)
Autonomous Mobile Robots (AMRs) are self-loathing robotic systems built to operate without human intervention in warehouse logistics environments. They rely entirely on onboard sensors, real-time processing, and AI to interpret the environment and make autonomous decisions, surpassing the traditional “one big CPU” mindset.
For decades, AMR brains have followed a familiar pattern. Route all sensor data to one powerful central processor, including camera feeds, laser range measurements, and inertial measurements. Its processor handles everything from slams (simultaneous localization and mapping) to building a map of the robot's environment, while also estimating its own location within that map, avoiding obstacles to motion controls.
This model worked well with lab prototypes and early deployments, but as AMRS expanded into the fleet and advanced into the real world, a single CPU approach began to show limitations. High delays, inefficient power usage, and calculation bottlenecks are all symptoms of intensive design that try to do too much.
Today's AMR navigates unpredictable warehouse aisles, adapts to real-time sensor feedback, and performs machine learning in-place inference. This means that developers cannot pay for inefficient systems. This is why developers are moving from centralized architectures to distributed computing architectures that push perception and control functions closer to the sensor itself.
Move from centralized architecture
In a centralized model, real-time tasks must compete for processing time on the same CPU. This not only increases latency, but also reduces determinism. This is a problem with tasks such as motor control, which requires sub-millisecond responsiveness.
Scaling is also inefficient. Double the number of sensors or actuators often means double the calculated load at one location. An over-spection of the worst-case scenario results in poor performance while the system's cost and power usage is inflated. These trade-offs are particularly severe on battery-powered platforms where all watts are counted.
Furthermore, centralized designs are difficult to modularize. Adding new sensors or updating subsystems often requires rebuilding the central firmware stack and recalibrating the computing resources.
The rise of edge intelligence
It employs a distributed computing architecture to offload preprocessing and inference tasks to embedded processors or microcontrollers within subsystems such as visual acuity modules, LIDAR arrays, and motor controllers.
This evolution is not just about replacing CPUs. This is a fundamental rethink of a system architecture that more closely aligns hardware and software with real-world requirements. Perception is just one area where you can benefit. Traditional AMR can stream high-resolution RGB or depth images to a central processor, potentially performing object recognition or slam algorithms. This involves a large amount of data transmission costs and delays.
Edge Computation Nodes can extract functionality, estimate depth, and even locally handle AI inference, providing compact semantic data rather than raw images. This shift means that the central processor doesn't have to be a power-hungry beast. Although we can focus on high-level planning and coordination, edge nodes perform time-critical and localized operations. result? It responds faster, reduces energy consumption, and reduces heat constraints.
Distributed computing
Distributed architectures are built around edge autonomy. In a typical setup, the AMR's Vision subsystem might include a stereo camera with an onboard processor that performs depth estimation and object detection locally. LIDAR (light detection and range) systems may preprocess the point cloud before sharing it with the navigation engine. The Inertial Measurement Unit (IMUS) feeds real-time pause updates directly to control the loop and bypasses the central path completely.
Key workloads such as SLAM, navigation planning, and obstacle avoidance still require integration, but are already increasing the amount of data processed. This reduces central processing needs and simplifies integration across diverse hardware configurations.
Microcontrollers also play a bigger role. For example, motor control requires that closed loop feedback be handled with microsecond level determinism. By embedding real-time controls in dedicated MCUs near the drive system, developers often use industrial protocols such as EtherCat to obtain close motion accuracy without loading the main processor.
Benefits beyond performance
Distributed computing offers several advantages beyond speed and responsiveness. There is one power efficiency. By activating only the processors needed for the task and allowing idle units to sleep, AMR can significantly extend operational uptime and promote charging infrastructure, maintenance cycles, and total cost of ownership.
Reliability is another advantage. A subsystem with local intelligence can continue to operate even if the central processor is busy or defective. For example, a battery management system with its own MCU can protect the battery pack during software crashes elsewhere. Similarly, safe features such as collision avoidance can be isolated.
Distributed architectures also enable modularity. Developers can replace subsystems with minimal impact on the rest of the stack. This accelerates iteration, simplifies inter-platform reuse, and allows AMR to tailor it to your niche environment and tasks. Modularity also helps build more complex systems on top of AMR platforms, such as dual-arm robotic systems.
Software Challenge
Of course, edge intelligence is not without costs, but it brings the complexity of integration. Task distribution refers to synchronizing multiple processors, managing interconnect latency, and maintaining software consistency across uneven hardware.
Here, a layer of abstraction is provided for middleware such as ROS 2 (robot operating system). The Publish-Subscribe model of ROS 2 natively supports distributed systems, allowing sensor nodes, control units and AI modules to communicate without tightly combining implementations.
To make everything work, the software stack must be co-designed with hardware constraints in mind. Developers should optimize resource-limited MCU or NPU inference models, manage thermal budgets carefully, and ensure that delay-sensitive tasks remain critical.
Laying the foundation for more efficient robots
Designed to complete a huge range of tasks in a variety of environments, AMR, of course, doesn't have an answer for all sizes. Centralized designs still play a role in deploying low complexity AMR or constrained costs. A robot with minimal sensing or pre-programmed routes may benefit, for example, from a single, integrated computing platform.
But for scalable and responsive AMRs that run in dynamic environments, the future is definitely on the edge. Engineers need to weigh the design trade-offs between latency, power, scalability, and integration complexity, which are choices that are likely to be shaped by applications and scale.
What's clear is that yesterday's computing architecture, even though its simplicity is elegant, no longer meets the demands of modern autonomy. From perception to actuation, distributed intelligence is becoming the foundation of agile, efficient, mission-ready mobile robots.
Dr.-ing. Nicholas Lement Leading the NXP robot team and overseeing the Industrial Systems Innovation Board. Before joining NXP, he designed cutting-edge computer vision and robotic systems for ABB and Smartray. He collaborated in his research paper on topics ranging from ML-driven video classification around human pose tracking to collaborative robotics. This academic study earned him a PhD from Munken University.
