When TOPS Is Not the Only Deciding Spec: A Robotics Platform Built for AI, Vision, and Control

Advantech’s robotics platform is built around the Qualcomm Dragonwing™ IQ-9075 processor to address these system-level requirements. The ASR-A503 robot controller SBC and the AFE-A503 industrial robot controller system use the Dragonwing IQ-9075M, the module form factor of the IQ-9075, to support robotics teams with a unified controller architecture designed for AI inference, multi-camera vision, robot-facing I/O, and deployment flexibility.

The Robotics Workload Is Not One Workload

A robot controller is not a single-purpose AI inference device. It has to run multiple workload classes at the same time, each with different performance and latency requirements.

Perception: Perception requires camera input, image pre-processing, AI inference, LiDAR and IMU data handling, and sensor fusion. These workloads are throughput-sensitive and benefit from dedicated AI and vision processing.

Planning: Planning requires CPU resources, memory headroom, and the ability to handle changing robot state. These workloads are logic-heavy and can require sudden bursts of compute during navigation, localization, and path updates.

Control and I/O: Control and I/O require predictable response across CAN FD, DIO, Ethernet, USB, serial, and other robot-facing interfaces. These workloads are not defined by peak AI performance, but by timing, reliability, and integration quality.

This is why robotics platforms should not be evaluated only by peak TOPS. TOPS measures AI throughput, but it does not answer the deployment question: can this platform support all three workload classes, perception, planning, and control, within the real constraints of a production system? The real choke points are memory bandwidth, power and thermals, and usable inference.  Engineering teams also need to evaluate performance per watt, camera support, I/O availability, memory capacity, thermal behavior, mechanical fit, operating system support, and the ability to move from prototype hardware to deployable systems.

The Advantech Platform: One Dragonwing IQ9 Architecture, Two Robotics Form Factors

Advantech addresses the requirement for a single controller platform that handles perception, planning, and control within real production constraints with two products built on the same Dragonwing IQ-9075M architecture.  Advantech selected Dragonwing IQ9 because robotics requires balanced compute across AI, vision, I/O, and control. Dragonwing IQ9 helps provide ASR-A503 and AFE-A503 with a heterogeneous architecture for distributing workloads across CPU, GPU, NPU, and vision processing resources.

ASR-A503: Board-Level Integration for Robot Designs

The ASR-A503 is the robot controller single board computer (SBC) for board-level integration. The ASR-A503 is built for teams that require a compact robotics controller inside a custom system design. It uses the Dragonwing IQ-9075M and supports up to 100 dense TOPS and 200 sparse TOPS AI performance for on-device perception workloads.

Perception workloads can use dedicated AI and vision acceleration. Planning and application logic can run on the CPU. Camera and video processing can use the integrated vision pipeline. Robot-facing I/O is handled through Advantech’s board and system design, including GbE, USB, CAN FD, DIO, serial, M.2 expansion, and optional GMSL.

AFE-A503: Industrial-Grade Robotic Controller for Quick Deployments

The AFE-A503 converts the ASR-A503 robotics controller architecture into an edge AI box form factor. It is designed for AMR control systems and robotics deployments that require a fully integrated controller with housing, power input, isolated interfaces, and system-level integration.

Where Real Edge Systems Fail Fast  

GPU-centric platforms can be strong for AI throughput, and they remain useful for many edge AI workloads. Robotics systems, however, need to be evaluated at the system level. Runtime, thermal design, battery size, cooling strategy, and enclosure constraints are all affected by performance per watt.

In benchmark comparisons against a competitor GPU-centric edge platform, the Dragonwing IQ-9075M-based platform showed stronger inference efficiency across representative vision workloads. 

These efficiency differences matter because the perception load runs continuously while the robot is operating. Drawing more power for the same work increases heat output and raises the demands on the thermal design and battery, which in turn reduces operating time between charges. For a battery-powered robot, that made the higher-throughput GPU platform the less practical option, even though it performed well on the benchmark itself.

Advantech weighed several other factors alongside efficiency, again favoring system fit over just the TOPS specification:

• Multi-camera vision: Supports 16 camera inputs, exposed as 8x GMSL plus MIPI-CSI lanes, into an integrated vision and AI pipeline. Perception around the body, not a single forward camera.
• AI and control on one SoC: Perception, sensor fusion, planning, and real-time control on one heterogeneous platform, not an inference accelerator that still needs a motion board beside it.
• Robot-facing I/O: Isolated CAN-FD for motors, isolated 16-bit DIO for e-stops and indicators, 4x GbE with PoE, M.2 expansion, 20 to 36 VDC straight off the battery rail.
• Production hardening: ECC memory, -40 to 85°C operating range, ESD and surge protection at the pins, TPM 2.0, lifecycle support to 2038.

Summary

Controller decisions are not settled by raw TOPS, but by the platform that can run the robot's full workload within real-world constraints of power, thermal design, sensor integration, I/O, and deployment readiness.

The ASR-A503 and AFE-A503 aim to provide robotics teams one architecture for the complete robot workload, not AI inference alone. The ASR-A503 supports board-level integration for custom designs, while the AFE-A503 is developed as an industrial controller system for faster deployments. Both are based on the Dragonwing IQ-9075M and support local AI inference, multi-camera vision, robot-facing I/O, and Advantech Robotic Suite.