Physical AI sounds grand, but in a factory it comes down to something quite practical: can the AI see or sense what is happening, make a useful decision and connect that decision back to the real operation? If it ends at another dashboard or demo, it has not gone very far.
The market signals are becoming clearer. In its latest first-half 2026 investment update, the Malaysian Investment Development Authority (MIDA) reported RM218.5 billion in approved investments and highlighted AI-enabled automation, smart manufacturing and higher-value semiconductor manufacturing as part of Malaysia’s industrial direction. MIDA also cited locally developed AI-embedded precision automation as an example of the capability Malaysia is trying to build.
At the same time, the international industrial ecosystem is shifting from AI used mainly for dashboards and analytics towards systems that can perceive, decide and influence the physical process. The Edge AI and Vision Alliance recently described vision-language models and physical AI as moving from promising concepts into real systems, while manufacturing technology providers are increasingly combining computer vision, edge inference, robotics, metrology, digital twins and condition monitoring.
What does Physical AI mean in a factory?
Physical AI does not simply mean a humanoid robot. In manufacturing, it is better understood as an AI system that is connected to the physical world through sensors, cameras, machines, actuators, robots and industrial control systems.
A useful architecture is:
Sensors / cameras / machine data → edge processing → AI model → engineering rules and constraints → PLC / robot / MES / CMMS / operator action.
The important part is the last step. If an AI model detects a defect, predicts a bearing problem or identifies a production bottleneck but nobody acts on it, the system is still only an analytics tool. Physical AI becomes valuable when its output is integrated into the operational workflow.
1. AI machine vision is one of the best places to start
Machine vision has become one of the most practical entry points for industrial AI because the business case is usually measurable: defect escape, false rejection, inspection labour, cycle time and traceability can all be quantified.
Traditional rule-based vision remains excellent when the inspection condition is stable and the defect is geometrically well defined. AI becomes attractive when normal products vary, defects are difficult to describe with fixed thresholds, surfaces are irregular, or several defect classes must be recognised at once.
Recent industrial activity reinforces this direction. On 2 September 2026, Hexagon announced a new optical CMM platform that combines vision, multisensor inspection and phased AI-assisted measurement capabilities. The Association for Advancing Automation is also hosting a September 2026 session specifically on AI-powered machine vision running on industrial PCs, including deep OCR, anomaly detection and edge inference.
For a factory, the most sensible first project is normally a single inspection station with a clearly defined defect cost. Prove detection performance, false reject rate, cycle time and maintainability before scaling across the line.
2. Predictive maintenance should move beyond dashboards
Vibration, temperature, motor current, acoustic and process data can identify developing equipment problems long before a conventional alarm threshold is reached. But many predictive-maintenance projects stall because the AI output remains separate from the maintenance workflow.
A more complete implementation connects condition monitoring to:
- asset identification and equipment hierarchy;
- fault classification and confidence;
- remaining risk or degradation trend;
- maintenance priority;
- CMMS work-order generation;
- spare-parts availability;
- maintenance feedback after inspection or repair.
This is where AI becomes operational rather than decorative. An early bearing warning is useful; an early bearing warning linked to the correct motor, maintenance history, spares and production schedule is much more valuable.
3. Autonomous material movement is becoming more realistic
AGVs and AMRs are not new, but physical AI is changing how autonomous equipment deals with variation. Instead of depending entirely on fixed routes and highly structured environments, newer systems increasingly combine computer vision, LiDAR, mapping, localisation and remote exception handling.
The practical opportunity for Malaysian manufacturers is not necessarily a fully autonomous warehouse on day one. A lower-risk starting point may be movement between two repetitive production locations, such as:
- raw material store to machine;
- machine to inspection;
- finished goods to staging area;
- tool or fixture delivery;
- automatic pallet or trolley transfer.
The engineering challenge is integration. The autonomous vehicle needs to know not only where to go, but when the machine is ready, what material is required, whether the route is available, and what to do when the process changes. That requires communication with PLCs, WMS, MES or production scheduling systems.
4. Adaptive process control is the next step after prediction
Many factories already use AI to predict quality, energy consumption or process performance. The next step is to let the optimisation layer recommend — and eventually adjust — process setpoints within defined engineering limits.
Examples include:
- optimising HVAC or chilled-water operation against load and energy cost;
- adjusting process parameters to maintain product quality;
- balancing machine throughput against energy consumption;
- optimising production schedules against actual equipment availability;
- using digital twins to test candidate operating strategies before applying them.
This should not be implemented as unconstrained AI control. Industrial systems need deterministic safety limits, operating envelopes, interlocks, fallback modes and auditability. A good architecture keeps the PLC or safety system responsible for deterministic protection while the AI layer performs supervisory optimisation.
5. AI agents can automate the workflow around the machine
One of the more important 2026 developments is the move from predictive models to AI agents that can coordinate several software systems. In maintenance, for example, the sequence could become:
Detect abnormal vibration → identify likely fault → check maintenance history → examine planned production → check spare stock → propose maintenance window → create CMMS work order → notify responsible engineer → learn from the repair outcome.
This is more valuable than simply adding a chatbot to a factory dashboard. The AI agent needs controlled access to real operational systems, clear permissions and traceable actions.
What I would start with
| Use case | Typical first KPI | Integration difficulty |
|---|---|---|
| AI vision inspection | Defect detection / false reject / inspection cycle | Low to medium |
| Predictive maintenance | Avoided downtime / warning lead time | Medium |
| Autonomous material movement | Trips per hour / labour / waiting time | Medium to high |
| Adaptive process optimisation | Energy / yield / throughput | High |
| AI maintenance or production agent | Response time / administrative effort / schedule adherence | Medium to high |
Do not start with the most impressive demo
Factories should resist the temptation to begin with the most impressive AI demonstration. Start with the process where four things are already true:
- the problem is expensive enough to matter;
- the required data can be obtained reliably;
- the output can be connected to an operational decision;
- the result can be measured against a baseline.
The technical implementation can then be expanded incrementally. Brownfield machines do not need to be replaced. Existing PLCs, SCADA, OPC DA, OPC UA, Modbus, industrial Ethernet, MQTT and database systems can often be connected through an integration layer, allowing AI functions to be added without disturbing deterministic machine control.
What interests me more is the move from AI that only observes the factory to AI that actually participates in the workflow. That part is less about fashionable models and much more about industrial integration: signals, permissions, control boundaries and what happens after the AI makes a decision.
CANS works across embedded systems, industrial networks, machine data acquisition, edge computing, AI analytics, digitalisation, machine vision, condition monitoring, SCADA/MES integration and digital-twin applications. This makes it possible to start from the physical machine and connect progressively upwards instead of creating an isolated AI application.
If you are considering AI vision inspection, predictive maintenance, autonomous scheduling, intelligent material movement or closed-loop process optimisation, send us one current production problem. We can help determine whether AI is justified, what data is required, how it should connect to the existing plant, and what a realistic proof-of-concept should measure.
Discuss the application with CANS on WhatsApp or contact us here.
Recent Comments