Where AI Actually Sits in an Industrial Robot Cell
The industrial robot arm itself — the FANUC, ABB, KUKA, or Yaskawa Motoman on the floor — runs deterministic motion control from a native controller (KRC5, IRC5, R-30iB, YRC1000). That code is not AI. It is inverse kinematics, PID loops, and taught points executed cycle after cycle to a repeatability specified in the datasheet.
AI enters through a second box wired to that controller: a vision system, a perception PC, or a cloud service. The International Federation of Robotics 2024 trends brief frames AI in industrial robotics as bin picking, quality inspection, and predictive maintenance rather than a replacement for the motion controller. The Bot Scout view is the same: buy the arm on cycle time and reliability, then price the AI module against the process problem it is supposed to solve.
The uncomfortable pattern is that adding AI without fixing part presentation, lighting, or fixture repeatability moves the failure mode rather than removing it. A neural net that classifies 'unknown' 30 percent of the time is not an AI problem, it is a lighting problem.
| Cell function | Classical control | Where AI is used today | Named example |
|---|---|---|---|
| Point-to-point motion | Taught path, inverse kinematics | Rarely — motion stays deterministic | Native controllers (FANUC R-30iB, ABB IRC5) |
| Bin picking (random parts) | Fails without structured feed | 6-DoF pose estimation on point clouds | Mujin MujinController, Berkshire Grey RIOS |
| Quality inspection | Rule-based vision, thresholds | Deep-learning defect classification | Cognex ViDi / Vision Pro Deep Learning, Landing AI |
| Predictive maintenance | Fixed service intervals | Anomaly detection on current, vibration, torque | FANUC ZDT, ABB Ability Condition Monitoring |
| Collision avoidance (dynamic) | Safety-rated stop, light curtains | Real-time replanning around detected objects | Realtime Robotics, Veo Robotics FreeMove |
| Simulation and training | CAD-based offline programming | Synthetic data + sim-to-real reinforcement learning | NVIDIA Isaac Sim, Isaac Manipulator |
Bin Picking: The Clearest AI Win
Bin picking is the application AI unlocked. Classical vision handles parts in a fixture; unstructured parts in a tote defeat it. AI perception stacks estimate 6-DoF pose from RGB-D or structured-light point clouds, then hand a pick pose to the robot controller. Mujin markets its MujinController as a motion-planning and perception layer that drives arms from FANUC, ABB, Yaskawa, and Denso against random parts in a bin. Berkshire Grey deploys the Robotic Induction and Order Systems (RIOS) family into e-commerce fulfillment with vision-guided picking. Ambi Robotics and Kindred (acquired by Ocado) sell into the same category.
Vision hardware under those stacks is dominated by industrial 3D cameras — Zivid, Photoneo MotionCam-3D, and Intel RealSense for lower-cost work. Grasping decisions come from learned models trained on synthetic data plus real captures. The honest metric is picks-per-hour on your parts, in your tote, at your tote height, with your gripper — not a demo video.
The buyer question is not 'does it use AI' but 'what is the pick rate and the miss rate on my SKUs, and what happens on a miss.' A cell that halts on every unknown item is not an AI cell, it is a very expensive light curtain.
Vision Inspection and Predictive Maintenance
Quality inspection is the second common AI slot. Cognex ViDi (part of the Vision Pro Deep Learning suite) and Landing AI LandingLens both sell deep-learning tools that classify defects a rule-based vision job cannot describe — surface scratches, weave variation, cosmetic flaws. Keyence, Zebra Matrox, and MVTec HALCON compete in the same space. The point is that a human labels a few hundred good and bad examples, and the model generalizes; the point is not that the robot becomes intelligent.
Predictive maintenance is the third. FANUC Zero Down Time (ZDT) collects controller data across a fleet and flags components trending toward failure. ABB Connected Services and KUKA iiQoT occupy the same slot. The service is anomaly detection on torque, current, temperature, and vibration — not general reasoning.
Collision avoidance is the fourth. Realtime Robotics ships a motion-planning accelerator that replans paths around dynamic obstacles in milliseconds; Veo Robotics (now part of Symbotic) sold FreeMove as a 3D safeguarded-stop system before winding down the standalone product. These are perception + planning modules on top of a safety-rated stop, not a replacement for ISO 10218-compliant guarding.
- Label the AI job precisely — 'find scratches longer than 2 mm on the machined face' beats 'find defects.'
- Insist on a pilot on your parts and your lighting before signing an inspection deal.
- For bin picking, measure picks per hour AND the miss handling — what does the cell do on an unknown object.
- Treat predictive maintenance data as an OT integration project (OPC UA, MQTT), not a subscription.
- Keep safety in classical, certified hardware; AI can suggest a stop, only a safety-rated system may enforce it.
ROS 2 and Isaac Sim: The Middleware Question
Do industrial robots use ROS? The answer is nuanced. Native controllers from FANUC, ABB, KUKA, and Yaskawa run proprietary real-time operating systems, not ROS. What has grown is ROS 2 (with ros2_control and MoveIt 2) as an integration layer — the PC that runs the vision stack, the fleet manager, or the mobile-manipulator base. The ROS-Industrial consortium at Southwest Research Institute maintains drivers and MoveIt configurations for many industrial arms; the arm still executes on the OEM controller, and ROS 2 sits on the perception/coordination PC.
On the simulation side, NVIDIA Isaac Sim (built on Omniverse) and the newer Isaac Manipulator workflow have become the reference tools for generating synthetic training data, testing motion policies, and doing sim-to-real reinforcement learning for grasping. Isaac Lab (formerly Isaac Gym) is the training environment; the models transfer to the real cell through ROS 2 or vendor SDKs.
The honest buyer test: if a vendor pitches 'AI-native' or 'ROS-native,' ask which controller executes the motion and which box runs the model. If both live in the same PC and both are the vendor's, plan for a longer integration and a smaller pool of service technicians.
What to Ask Before You Buy an 'AI' Industrial Robot
The Bot Scout script for AI-tagged industrial robot quotes is short. First, name the AI job in one sentence — the specific perception, classification, or planning task the model performs. Second, ask what happens when the model is wrong: stop, retry, escalate to a human, or route to a reject bin. Third, ask who owns the training data and the model weights after the project ends.
Fourth, ask which safety standards the cell is assessed against — ISO 10218-1:2025 for industrial robots, ISO/TS 15066:2016 for collaborative operation, and the machinery directive in your jurisdiction. AI decisions do not replace safety-rated hardware. Fifth, price the second-year cost of re-labeling and re-training as the product mix drifts — this is the line most quotes omit.
The point of the exercise is not to reject AI. It is to make the AI portion of the cell a specified, tested, priced module rather than a marketing adjective.
Bottom Line
Industrial robots are becoming AI-enabled in narrow, well-defined places — bin picking, defect inspection, predictive maintenance, and dynamic collision avoidance — while the motion controller itself remains classical. Buy the arm on the motion job and the AI module on the perception job, and hold each to its own metric.
Before signing an 'AI robot' quote, write the perception task in one sentence and demand a pilot on your parts.
FAQs
Do industrial robots use AI?
The motion controller on a standard industrial arm from FANUC, ABB, KUKA, or Yaskawa is deterministic control code, not AI. AI is added as a separate module — vision for bin picking, deep learning for defect inspection, anomaly detection for predictive maintenance — that hands results to the robot controller.
Do industrial robots run ROS?
The native controllers do not run ROS. ROS 2, with ros2_control and MoveIt 2, is widely used on the perception PC, the fleet manager, or the mobile-manipulator base — and the ROS-Industrial consortium maintains drivers for many industrial arms — but the OEM controller still executes the motion.
What is the clearest AI application for industrial robots today?
Bin picking of random parts. Vendors including Mujin, Berkshire Grey (RIOS), Ambi Robotics, and Kindred use learned pose estimation on 3D camera data — Zivid, Photoneo, RealSense — to pick unstructured items that classical vision cannot handle.
Where does NVIDIA Isaac Sim fit into an industrial robot project?
Isaac Sim, built on Omniverse, is used to generate synthetic training data for perception models and to develop and test motion policies before deployment. Isaac Manipulator packages the workflow for robot arms, and the trained models transfer to the real cell through ROS 2 or vendor SDKs.
Can AI replace safety-rated guarding on an industrial cell?
No. ISO 10218-1:2025 and ISO/TS 15066:2016 require safety-rated stop functions that AI perception cannot legally provide by itself. AI systems for collision avoidance from vendors like Realtime Robotics augment a certified safety-stop; they do not replace it.