AI robotics

AI robotics means robots steered by an AI model instead of a fixed program. A classic factory robot repeats programmed movements. An AI-driven robot looks through its cameras, takes an instruction such as "put the apple in the bowl", and works out the movements itself. For a business, the promise is a robot that switches tasks without new programming. The examples below are research models.

How does it differ from a classic robot?

In the ISO definition that the International Federation of Robotics uses, an industrial robot is an automatically controlled, reprogrammable manipulator: someone writes its motions for each task. AI robotics replaces that program with a model that has learned from examples. The wider field of AI that acts through a body is called embodied AI.

Much current work uses vision-language-action models (VLA), a name from Google DeepMind's RT-2 (2023). RT-2 takes a vision-language model trained on web images and text and trains it further on robot data and web data together, with the robot's actions written as text tokens: it answers an instruction with movements the way it answers a question with words. Across 6,000 test runs it coped better with unfamiliar objects than earlier robot models, and it followed commands that were not in its robot data, such as placing an object on a particular number or icon.

Where does the training data come from?

Robot data "scales linearly with human labor", NVIDIA writes in its GR00T N1 paper (2025), because a person usually steers the robot remotely for every demonstration, which the model then learns to imitate. GR00T N1, an open foundation model for humanoid robots, stacks its data as a pyramid: web data and videos of people at the base, synthetic data from simulations and video models in the middle, real robot data at the top.

A simulator also allows reinforcement learning, learning by trial and error, which on real hardware can be dangerous and take thousands of hours (Tobin et al., OpenAI and UC Berkeley, 2017). The catch is the reality gap: simulators miss effects such as wear and flexible materials. Tobin et al. randomised colours, textures and lighting so that the real world might look like one more variation; their object detector, trained only on simulated images, located real objects to within 1.5 cm.

Hugging Face counted 1,145 robotics datasets on its platform in 2024 and 26,991 in 2025, which made robotics its largest dataset category (March 2026).

How does the model steer the robot?

The model runs in a loop: look, act, look again. The biggest RT-2 model, with 55 billion parameters, ran in the cloud at 1 to 3 steps per second; its authors warn that computing cost may become a bottleneck for tasks that need faster control. GR00T N1 splits the work: a vision-language part reads the camera image and the instruction 10 times per second, and an action module sends motor commands 120 times per second.

Where does it go wrong?

• Web knowledge does not teach new movements: RT-2's physical skills stayed limited to those in its robot data, used in new ways.
• Every robot is built differently. The GR00T N1 paper speaks of "data islands": differences in body, sensors, joints and control keep the data of different robots apart.
• Simulators miss things. For surgery, NVIDIA names soft tissue, reflections, blood and smoke.

Example

In March 2026 NVIDIA presented GR00T-H, a VLA for surgical robots derived from its GR00T N models. It learned from about 600 hours of Open-H-Embodiment, an open dataset of 778 hours, mostly from surgical robots, built by 35 organisations. Each robot type gets its own small network that maps its movements onto a shared action space, so one model can learn from all of them. A prototype completed a full suture, from picking up the needle to tying the knot, in the SutureBot benchmark on a research robot.