Robotics gives a machine a body; AI gives it the ability to see, decide and learn. In 2026 the two are merging through robot foundation models, which let one model control many robots and tasks from camera input and plain-language instructions. Most of that AI now runs on the robot itself, on edge computers, not in the cloud.
Updated October 2026
What changed in 2026
- The installed base hit a record. About 5 million industrial robots now work in factories worldwide, up 9%. Over 600,000 were installed in 2025 (IFR, Sept 2026).
- Robot foundation models became usable. NVIDIA's open Isaac GR00T model reached version N1.7. Google DeepMind released Gemini Robotics 1.5 (Sept 2025) and Gemini Robotics On-Device 2 (July 2026, trusted testers only).
- Edge compute caught up. NVIDIA Jetson Thor went on sale in August 2025, with up to 7.5x the AI compute of Jetson Orin in a 130 W envelope.
- Humanoids moved from demos to paid pilots. Hyundai said at CES 2026 that Boston Dynamics' Atlas will start work at its Georgia Metaplant in 2028 (InsideEVs).
- Rules got closer. The EU Machinery Regulation applies from 20 January 2027 and, for the first time, counts software as a safety component (Pilz).
AI vs robotics: what each one does
Robotics and AI are different fields that now overlap.
| Robotics | AI | AI-powered robot | |
|---|---|---|---|
| What it is | Machines that sense and move in the physical world | Software that finds patterns, predicts and decides | A robot whose behaviour comes from learned models |
| How it is set up | Programmed step by step | Trained on data | Trained, then fine-tuned per task |
| Handles change? | Poorly; a new part means new code | Yes, within its training | Yes, within limits; new tasks need data |
| Example | Welding arm on a fixed path | Defect detection in images | Arm that picks unseen parts from a bin |
A traditional robot repeats a path very precisely. It fails when the part, the lighting or the position changes. AI closes that gap. It lets the robot see what is in front of it and choose what to do.
How AI is used in robotics
AI shows up in five main jobs inside a robot:
- Perception. Computer vision finds parts, people and obstacles. This is the oldest and most mature use. See machine vision applications.
- Planning and navigation. Mobile robots map a warehouse and route around people and forklifts.
- Manipulation. Learned grasping lets arms pick mixed, unseen items. Logistics is the leading sector for AI robots, according to the IFR position paper.
- Language and reasoning. Vision-language-action (VLA) models take an instruction such as "put the red parts in bin 3", look at the scene and output motor actions.
- Health and upkeep. Models read motor current, vibration and heat to flag failures early.
Robot foundation models, in plain English
A robot foundation model is one large model trained on many robots and tasks. It is the robotics version of a large language model.
The most common type is the VLA model. It combines a vision-language model with an "action head" that outputs joint movements. NVIDIA's GR00T works this way: a vision-language backbone plus a diffusion transformer that generates continuous actions (GitHub). Its code is Apache 2.0 and its weights use the NVIDIA Open Model License.
Why it matters for buyers:
- Less hand-coding. You fine-tune on demonstrations instead of writing a program for each task.
- Reuse across robots. One model family can serve different arms and humanoids.
- Still narrow in practice. Google DeepMind notes its on-device model generalises less well to unfamiliar tasks and to robots with many joints (model card). Plan for fine-tuning and testing on your own tasks.
Humanoid robots: where they really stand
Humanoids get the headlines. The facts in 2026 are more modest, but real:
| Company | Robot | Where | Status (as reported) |
|---|---|---|---|
| Hyundai / Boston Dynamics | Atlas | Metaplant America, Georgia | Parts sequencing from 2028, assembly by 2030 (target) |
| BMW / Figure | Figure 03 | Spartanburg, US | Logistics sequencing deployment announced June 2026 |
| GXO / Agility Robotics | Digit | Fulfilment centre | Over 100,000 totes moved (Agility) |
| Toyota Canada / Agility | Digit | Vehicle plant | Commercial agreement, Feb 2026 |
| Schaeffler / Hexagon | AEON | Schaeffler factories | At least 1,000 robots over seven years (plan) |
Source: The AI Insider, Aug 2026, which compiles company announcements.
The same report cites about 19,100 humanoid shipments worldwide in the first half of 2026 (Smart Analytics Global). That is small next to 600,000 industrial robots installed in 2025. Most value from AI in robotics today still comes from arms, mobile robots and vision systems, not humanoids.
Robotics and AI in manufacturing: mass customisation
The original version of this post focused on mass customisation: making products to each customer's specification at mass-production cost. That case is stronger now.
- Flexible cells. AI vision lets one robot cell handle many product variants without re-programming.
- Collaborative robots. Cobots work next to people on short runs and changeovers.
- Digital twins. Teams test a new variant or line layout in simulation before touching the floor. See how digital twin technology works.
- Inline quality. Vision models check each unit, so a custom order does not mean manual inspection.
How to deploy AI robots: a practical sequence
- Pick one task with a clear measure. Picks per hour, defect rate or downtime. Avoid "general automation" as a goal.
- Check the data you have. Camera footage, demonstrations and sensor logs. VLA models need task examples to fine-tune.
- Train and test in simulation first. Simulation is cheaper and safer than the floor.
- Choose edge hardware for the model size. Robots need answers in milliseconds and cannot wait on a network.
- Run a supervised pilot. Keep a person in the loop. Log every decision the robot makes.
- Set up model operations. Version models, roll out updates in stages and roll back fast.
- Document safety. Under the EU Machinery Regulation, software can be a safety component from January 2027.
Challenges to plan for
- Safety and certification. Learned behaviour is harder to certify than a fixed program.
- Fleet updates. A model update to 200 robots is a release, not a file copy. It needs staging, monitoring and rollback.
- Data. Robot training data is scarce and costly to collect.
- Skills. Teams need people who understand both machines and models.
- Cost. A Jetson AGX Thor developer kit alone starts at $3,499 (NVIDIA). Budget for integration, which usually costs more than the hardware.
Where it's heading: 2026 → 2030
| Period | What it looks like |
|---|---|
| Around 2020 | Robots mostly programmed path by path. AI used mainly for vision inspection and some bin picking. Humanoids were research projects. |
| 2026 (now) | 5 million industrial robots in operation (IFR). Open VLA models (GR00T N1.7) and on-robot compute (Jetson Thor). Humanoids in paid pilots. |
| By 2028 (expected) | EU Machinery Regulation in force (from Jan 2027). Atlas doing parts sequencing at Hyundai's Georgia plant; Hyundai targets capacity for 30,000 robots a year (InsideEVs). |
| By 2030 (expected) | IFR expects about 806,000 industrial robot installations a year by 2029. Hyundai targets Atlas on component assembly by 2030. Fleet-wide model updates expected to be routine. |
Where NexaStack fits
AI robots are only as reliable as the models running on them. NexaStack deploys, updates and monitors AI models and agents across edge devices, including robot controllers and on-site GPU servers. Teams can stage a model update to a few robots, watch how it behaves, then roll it out or roll it back, with a record of which model made which decision. See What Is Physical AI? for the bigger picture.
Frequently asked questions
How is AI used in robotics?
AI gives robots perception, decision-making and learning. Vision models find objects and people. Planning models route mobile robots. Vision-language-action models turn camera input and instructions into movements. Predictive models flag failures before they happen.
What is the difference between AI and robotics?
Robotics is about physical machines that sense and move. AI is software that learns from data to predict or decide. A robot can run without AI on a fixed program. AI can run without a robot, for example in a chatbot. An AI-powered robot uses learned models to handle tasks that change.
Will AI take over robotics?
AI is becoming the main way robots are programmed, but it is not replacing the engineering. Motors, sensors, safety systems and certification still matter. The IFR expects AI to be widely adopted in robotics over the next five to ten years (IFR).
Are humanoid robots working in factories in 2026?
Yes, in pilots and early commercial deals, not at scale. Examples include Figure at BMW in the US, Agility's Digit at GXO, and planned Atlas work at Hyundai from 2028. First-half 2026 humanoid shipments were about 19,100 units, as reported by The AI Insider.
What is a robot foundation model?
It is one large AI model trained on data from many robots and tasks. Most are vision-language-action models: they take images and text and output robot actions. Examples include NVIDIA Isaac GR00T and Google DeepMind's Gemini Robotics.
Run AI robots you can trust in production
Bring one robot task and we will map how NexaStack stages, monitors and rolls back models across your fleet. Book a 30-minute robot fleet AI pilot call →
