A robotics foundation model is one AI model trained to control many kinds of robot, such as arms, wheeled robots and walking robots, instead of a separate program written for each machine. It is a foundation model for the physical world. It reads what the robot's cameras and sensors report, and decides how the robot's motors should move.
Most industrial robots today follow programs written point by point by a specialist. An engineer tells the arm exactly where to move, how far to close the gripper and when to let go. This works well when every part is the same and arrives in the same place, as on a car assembly line. It fails when the objects vary. In a warehouse, every parcel has a different size, shape and weight, and the parcels arrive jumbled together. Writing a program for every case is impossible.
A robotics foundation model learns instead. It is trained on huge amounts of data: videos of people handling objects, recordings of many different robots doing tasks, and millions of attempts made in computer simulation, where a virtual robot can practise far faster than a real one. Training runs on thousands of GPUs, often coordinated by a distributed computing framework; the page on AI accelerators covers these chips. The finished model takes in camera images and the position of each joint, many times a second, and outputs the next movement for each motor.
The aim of one model for many machines is that each new robot does not need years of new programming. A company adapts the model to a new arm or a new walking robot with a smaller amount of data from that machine.
Take, for example, a parcel sorting centre in Bhiwandi, near Mumbai, during a festive sale. A robot arm stands over a chute where parcels arrive mixed together: boxes, soft pouches and a cricket bat in bubble wrap. The arm has never handled this bat. The model looks at the camera image, judges the shape and chooses to grip the bat near its middle, where it will balance. It lifts the bat and places it on the belt for Pune. The arm handles about 600 parcels an hour this way. When it cannot find a safe grip, it stops and asks a person for help.
Such models are used for robotic picking in warehouses and for AI robotic picking by arms. Later they could control industrial robots and goods-to-person robots. The field is young. Today's models still make mistakes that a person would not, and every robot body still needs testing before it can work safely beside people.