What you should know:

• Agentic AI is increasingly running across platforms, orchestrating intelligence between smartphones to PCs and robots. 
• As intelligence moves across devices, the AI models running on those devices determine how AI agents optimize factories’ operations. 
• The convergence of industrial intelligence, edge computing and agentic AI abstracts the complexity of industrial machines, enabling agents to run them through natural language. 
• In this post, we explore how Forgis leveraged Qualcomm Technologies’ distributed architecture to bring foundation models to the factory floor, running on the most advanced Arduino edge compute platform. 

When agentic AI meets the industrial floor

Think about the impact of agentic AI on your day-to-day life. At work, assembling data today just takes a good prompt and a few back-and-forth messages with your AI agent. At home, your agent is a few words away from organizing your photos or finding the best tummy ache relief.

You focus on crafting a good prompt and let your AI agent execute. The technology under the hood — foundation models, orchestrators, compute power and memory, tool calling, APIs and more — removes the complexity of interfaces. This is exactly what is now happening in industrial settings.

As AI becomes the new UI, natural language is removing the physical barriers that would slow down the benefits of AI and automation, leaving room for plants that are adaptive, self-optimizing and capable of responding in real time. The factory operator who once needed specialized training to operate a complex machine can now just tell it what to do, in their own words.

Bringing intelligence to industrial AI

Swiss startup Forgis, which builds physical AI models for manufacturing, collaborated with Arduino, a Qualcomm Technologies company, to let an operator control a robotic arm using their voice.

Here is what the experience looks like in practice:

• The operator uses a smartphone to send voice commands to an AI agent running on the Arduino UNO Q board.
• The Forgis foundation model processes the prompt and determines which object to pick, where to place it and calculates the full motion plan.
• The robot completes the task, placing different items into boxes.
• UNO Q uses its LED matrix to display the agent’s current state and summarize the action it’s taking in real time.

The technology

In this example, Forgis’ proprietary foundation model receives multimodal factory data — robot’s CAD model, PLC I/O signals, project specifications — and translates them into structured skills, parameters, and contextual information for downstream robotic control and machine execution.

Several technologies work together to make this happen. On the hardware side, a Samsung smartphone powered by Snapdragon handles natural language processing, sending the transcript of the prompt directly to the UNO Q, powered by a Qualcomm Dragonwing QRB2210 processor.

Automatic speech recognition and transcription capture the operator’s instructions, while the Forgis foundation model converts a prompt like “put each box in its respective compartment” into a structured set of machine-readable instructions:

• Connect to robot camera
• Identify objects
• Calculate pick-and-place sequence
• Execute motion plan

The robotic arm’s camera is connected via USB to send the video stream directly to the board, which runs the Forgis foundation model, able to recognize the number and the position of the expected components with a latency of 20 milliseconds. The model output is used to determine the pick-and-place sequence and execute the motion plan.

Watch this demonstration:

Benefits

The value is tangible for manufacturers and machine makers. AI agents are designed to reduce the time technicians spend on routine floor operations, freeing them up for higher-value tasks like oversight, quality control and exception handling.

Fewer manual inputs also mean fewer errors, which is a critical advantage in precision manufacturing environments like aerospace, medical devices and automotive. From there on, the foundation model continuously learns from production data, feeding insights to agents that make decisions in real time and optimize processes.

Orchestrating AI model inference entirely locally across the smartphone and UNO Q adds another advantage. A factory floor cannot wait for a cloud response — foundation models that run directly on the edge process data where it is generated, at the speed the process demands. No latency. No dependency on a network connection that may not exist. No production data leaving the facility. For industries where data sovereignty and uptime are non-negotiable, this is a meaningful differentiator.

Plus, thanks to the open Arduino ecosystem, the Forgis team can rapidly deploy their proprietary AI models without getting slowed down by hardware integration challenges. This is a testament to how accessible developing for edge and agentic AI in industrial settings has become.

What’s next?

Agentic AI is no longer confined to digital environments. It impacts the way we interact with our physical space in industrial settings — and even on our own laptop.

AI doesn’t just think; it acts in the physical world. Qualcomm Technologies is helping to power this shift by developing technologies designed to support intelligent systems that can connect voice commands with machine actions.

The Forgis implementation is just one example. As more developers build on platforms like UNO Q, expect to see agentic AI show up in places you’d never expect — and do things you’d never thought to ask for.

Go Deeper

What is the Qualcomm Dragonwing QRB2210, and what role does it play in this implementation?

The QRB2210 is the processor powering the Arduino UNO Q board at the heart of this demonstration — the compute engine that runs the Forgis foundation model directly on the factory floor, processing multimodal inputs including the robot's CAD model, PLC I/O signals and project specifications with a latency of just 20 milliseconds. It is what makes cloud-free, real-time inference possible in an industrial environment where every millisecond of responsiveness matters. 

The Forgis demo shows AI acting in the physical world — where does this go from here?

Agentic AI is no longer confined to digital environments: it is reshaping how we interact with our physical space in industrial settings, with intelligent systems now capable of connecting voice commands directly to machine actions.

Can developers and machine makers build on the same platform today?

The Forgis team built on the open Arduino ecosystem precisely because it allows them to rapidly deploy their proprietary AI models without getting slowed down by hardware integration challenges — and that same accessibility extends to any developer looking to bring agentic AI into physical environments.