History of AI in robotics: from specialized networks to generalist robots

Timeline: three eras of AI in robotics

Three eras follow one another, each adding a capability without erasing the last. The first invents programmed manipulation: in 1961, Unimate joins a General Motors assembly line, followed by Shakey, the first mobile robot able to plan its own actions, developed at SRI from 1966 to 1972. Industrial vision is born with players such as Cognex, founded in 1981, then 2D vision guides robots across the automotive industry throughout the 1990s and 2000s.

The second era brings deep learning into perception: AlexNet wins the ImageNet competition in 2012, YOLO publishes real-time object detection starting in 2015, Dex-Net extends deep learning to bin picking from 2017, and Covariant puts AI-powered warehouse picking into production in 2020.

The third era chases generalization: Google's RT-1 ships in 2022, followed in 2023 by RT-2 and by Meta's Segment Anything (SAM). Physical Intelligence's π0 arrives in 2024, then NVIDIA releases GR00T N1's open weights in March 2025. Through 2025 and 2026, the first humanoid pilots enter factories, such as the AEON deployment at Schaeffler or the GXO and Agility Robotics agreement for Digit.

AI is already here: a decade in production

Even before VLA models, AI has already been running in production in industrial robotics for more than a decade. Object detection with YOLO counts parts, spots PPE or sorts parcels. Image segmentation is moving fast with Meta's Segment Anything (SAM, April 2023), SAM 2 (July 2024) and SAM 3 (November 2025), which speed up annotation and picking. 2D and 3D vision-guided robotics, 6D pose estimation for bin picking (Photoneo, Zivid) and quality inspection through anomaly detection (Cognex, Keyence) are already running in thousands of factories.

Generalist warehouse picking is the most visible example. Amazon hired Covariant's founders and licensed its models in August 2024 to accelerate its robot fleet. In Germany, Sereact claims more than 200 systems deployed across Europe and over a billion production picks, for customers such as BMW and Mercedes-Benz. The message is simple: a task-specific convolutional network, integrated by a systems integrator, already works today, with no need to wait for a generalist model.

Specialized stack or generalist stack?

Both stories coexist today, on the same factory floor. Here is how the two stacks compare, brick by brick.

The specialized stack remains, today, the more reliable and easier to certify: every brick is tested, documented, independently replaceable, and its worst-case behaviour is predictable. Its cost is human: every new task demands fresh integration work. The generalist stack promises the opposite: a single model learned from demonstrations, able to adapt to new tasks without reprogramming. But in 2026 it remains overwhelmingly at the pilot stage, with open questions on reliability, certification and onboard inference cost. The two approaches are not mutually exclusive: many manufacturers already combine a specialized stack for critical tasks with generalist bricks for perception or high-level planning.

What this history teaches an industrial team

The current wave of physical AI is not a rupture out of nowhere: it is the third stage of an evolution more than sixty years long. Each stage reused the skills of the last rather than replacing them: the vision integrators who wired up factories in the 1990s are largely the same people deploying VLA pilots today, and production data quality remains the limiting factor now just as it was then.

Three reads to place your project within this continuity: our guide to VLA models to understand what the third era actually changes, our factory roadmap to sequence a pilot without betting everything on a still-young technology, and our industrial and mobile robots comparator to see how much of your current fleet already sits on the proven specialized stack.

Quick answers

What is YOLO? A family of neural networks that detects and localizes objects in an image in a single pass, in real time. Introduced in 2015, it remains the reference for counting, sorting or spotting objects on a factory floor.

What is SAM? The Segment Anything Model, developed by Meta from 2023 onward, precisely outlines the boundary of any object in an image or video without task-specific training. It speeds up data annotation and warehouse picking.

Why is everyone talking about VLAs now? Because vision-language-action models promise to replace several specialized bricks with a single model learned from demonstrations. It is promising but recent: most deployments remain pilots, not yet large-scale production.