micro-ROS vs ROS 2 on Embedded Linux

micro-ROS excels at bridging resource-constrained microcontrollers (MCUs) to the ROS 2 ecosystem because it implements a lightweight, real-time capable client library. For example, it runs on bare metal or an RTOS like FreeRTOS on an STM32F4 with as little as 128KB of RAM, achieving sub-millisecond latency for sensor-to-actuator loops via Micro XRCE-DDS. This makes it the de facto choice for deeply embedded sensor nodes and motor controllers where a full Linux kernel is impossible.

ROS 2 on Embedded Linux takes a different approach by running the full rclcpp or rclpy stack on a single-board computer like a Raspberry Pi 4 or NVIDIA Jetson. This results in a richer feature set, including direct support for complex nodes like SLAM, image processing, and standard DDS security plugins. The trade-off is a significantly larger memory footprint (typically >200MB of RAM) and a non-deterministic latency profile due to the Linux kernel scheduler, which can introduce jitter of 1-5ms in a standard PREEMPT_RT configuration.

The key trade-off: If your priority is hard real-time control on a sub-$10 BOM cost with minimal power draw, choose micro-ROS. If you prioritize algorithmic complexity, high-bandwidth sensor processing, and a unified development environment on a single board, choose ROS 2 on Embedded Linux. Consider micro-ROS for distributed sensing and actuation, and standard ROS 2 for the central compute node.

micro-ROS excels at bridging the gap to deeply resource-constrained microcontrollers because it implements a client-to-agent architecture that offloads complex middleware tasks. For example, a micro-ROS node on an STM32F4 (ARM Cortex-M4) can publish sensor data directly to a ROS 2 graph while consuming less than 50KB of RAM, a footprint impossible for a standard rclcpp node. This makes it the only viable choice for bare-metal or RTOS-based motor controllers and sensor hubs that require hard real-time determinism.

ROS 2 on Embedded Linux takes a different approach by running the full middleware stack on a more capable processor, such as a Cortex-A running Yocto-based Linux. This results in direct access to the complete ROS 2 ecosystem, including rclcpp, lifecycle nodes, and complex DDS configurations, but at a significantly higher resource cost—typically requiring 256MB+ of RAM and a full networking stack. The trade-off is raw capability and software composability versus power efficiency and BOM cost.

The key trade-off: If your priority is ultra-low-power sensor actuation with hard real-time constraints on a sub-$5 MCU, choose micro-ROS. If you prioritize running complex algorithms like SLAM or model predictive control directly on a rich embedded compute module, choose ROS 2 on Embedded Linux. In practice, a production robot will use both: micro-ROS for the sensorimotor edge and ROS 2 on Linux for the high-level compute brain, communicating seamlessly over a DDS-XRCE bridge.