A recent DIGITIMES column puts the spotlight on a transformation that is redefining the boundaries of industrial and service robotics. We are no longer talking about mechanical arms with add-on features, but a paradigm shift: so-called “Physical AI” moves the center of gravity from robots executing pre-programmed tasks to machines that perceive, reason, and act in real time. It’s a qualitative leap that involves not just algorithms, but the entire hardware supply chain and deployment choices.

Intelligence moves onboard

Until recently, a smart robot was often a cloud-connected entity: heavy computation was offloaded to remote servers, leaving sensors and actuators on board. Today the direction is the opposite. Zero-latency requirements, protection of intellectual property, and robustness in harsh or disconnected environments push toward fully local inference. It is no coincidence that major chipmakers are integrating Neural Processing Units (NPUs) and embedded GPUs with memory bandwidth tailored for computer vision and small language model workloads.

For robotics developers, the message is clear: hybrid computing gives way to extreme on-premise solutions, where every millisecond counts and processing happens entirely on the edge device. This means rethinking the hardware profile: no longer generic microcontrollers, but SoCs with dedicated accelerators and the ability to run quantized models — typically INT8 or FP16 — with context windows broad enough to handle complex instructions and operator-machine dialogue.

Constraints that dictate specifications

The transition to truly smart robots comes at a cost. Unlike the cloud, where resources can be over-provisioned, a mobile robot or a collaborative manipulator has limited thermal and energy budgets. This introduces trade-offs familiar to anyone working with Large Language Models locally: the choice of quantization level, model size, available VRAM, and memory bandwidth become critical variables. A system with only 8 GB of unified memory, for example, can run small language models with specific fine-tuning, but will struggle with complex vision-language models without aggressive compression techniques.

It’s a whole new TCO problem. The cost of the robot is no longer dominated by mechanics or gearmotors, but by onboard computational capability and the infrastructure for model updates and maintenance. At the same time, sovereignty of locally processed data becomes an architectural requirement: sectors like healthcare, defense, or critical logistics mandate that no raw data leave the machine, aligning with strict regulations. In this sense, physical AI is the battlefield for compliance by design.

From occasionally intelligent features to autonomous agent

This is not about adding “AI features” to a robot — object recognition, voice commands, navigation — but about building an autonomous agent that understands the operational context. This requires software pipelines that integrate perception, reasoning, and action in a single local flow. Orchestration frameworks like ROS 2 combined with edge-optimized inference runtimes become indispensable. And testing these pipelines can’t rely solely on simulation: a hardware-in-the-loop approach is needed to expose real bottlenecks, especially when compute power is below that of a server.

For those evaluating on-premise deployment in robotics, the analysis cannot stop at comparing GPU families or accelerators. The entire lifecycle must be considered: from training (often still done in the data center) to field inference, through model maintenance and over-the-air update management. At AI-RADAR we observe how these choices intertwine with the maturation of tools for model compression and distribution on embedded hardware — an ecosystem still fragmented but evolving rapidly.

A signal for industry and research

DIGITIMES’ contribution, even in its brevity, captures a turning point. It’s no longer science fiction: robots with local cognitive capabilities are entering factories and will soon appear in open environments. This forces designers and system integrators to upgrade their skills, embracing concepts like quantization, computational graph optimization, and multi-task model distribution on heterogeneous hardware. Otherwise, they risk sticking to an idea of “feature-based” automation that the market is rapidly leaving behind.