Robot foundation models

For fifty years, making a robot do something meant writing the something down. An engineer specified the positions, the speeds and the forces, tested until the failures stopped, and the resulting program worked on that robot, in that cell, for that part. Move the bench and it broke.

A robot foundation model is the opposite arrangement. One model is trained on an enormous mixture of robot demonstrations, human video and internet-scale image and text data, and it produces motor commands directly from what the robot sees and what it is told to do. It is not programmed for the task. It is adapted to the robot.

What VLA stands for, and why the order is the point

The family name is vision-language-action. The model takes camera images and an instruction in ordinary words, and outputs actions: joint positions, gripper commands, the small continuous numbers that actually move a machine.

Where a robot foundation model sits

The reason the field split into two models is timing. Reasoning about a ten-minute task is slow and can afford to be. Keeping a limb stable is fast and cannot wait for a thought. So the reasoning model plans, watches and calls the action model as a tool, and the action model runs the body at a much higher rate.

Who is building them

Google DeepMind publishes the Gemini Robotics family. Its July 2026 release runs to three models: an action model that controls a full humanoid from feet to fingertips, an embodied reasoning model that plans and orchestrates and is available through the Gemini API, and an on-device model built to run on the robot without a network. DeepMind reported driving three different robot bodies from a single model checkpoint.

Physical Intelligence works in the open more than most. It released π0 in October 2024 as its first generalist policy and open-sourced the weights and code in February 2025. π0.5 followed in April 2025 with open-world generalisation, demonstrated on a mobile manipulator tidying a kitchen it had never seen. π*0.6 in November 2025 added reinforcement learning from the model’s own experience, and π0.7 in April 2026 is described by the company as steerable, with a step change in generalisation.

NVIDIA publishes Isaac GR00T, an open foundation model family for humanoid reasoning and skills, trained on a mixture of real captured data, synthetic data and internet-scale video, and designed to be post-trained for a specific robot. NVIDIA’s angle is the whole stack: simulation for generating training data, and the on-robot compute to run the result.

Several robot makers train their own. Figure has Helix. The distinction that separates the two groups is whether the model is meant to run one company’s machine or anybody’s.

The transfer problem, which is the whole problem

A policy trained on one robot arm has historically been worthless on a different one. The joints differ, the reach differs, the camera sits somewhere else, and the model has learned the relationship between what it saw and what it did on that specific body.

This is why robotics has no equivalent of the internet-sized text corpus. Every new robot has meant a new dataset, collected by hand, mostly by teleoperation.

The current generation attacks it from two directions. Cross-embodiment training mixes data from many robot types so the model learns the task rather than the machine. Fast adaptation then tunes the trained model onto a new body with a small amount of data: DeepMind says its on-device model reaches a new bi-arm robot in a few hours, typically on fewer than 200 examples, across bodies with different shapes, sensors and degrees of freedom.

If that holds up outside the labs that publish it, the economics change. A robot maker stops needing a machine-learning team and starts needing a driver.

What these models are still bad at

Fine manipulation. DeepMind’s own July 2026 results put whole-body and gripper tasks at medium to high success and multi-finger dexterous manipulation at challenging. Everyone’s charts have the same shape.

Long tasks. Reasoning models now track jobs lasting minutes and hundreds of decisions, and the honest framing is that this is recent. Physical Intelligence published work in March 2026 on giving models long and short-term memory specifically to get past ten minutes.

Knowing they have failed. A model that has never seen a failure does not recognise one. Progress tracking, where DeepMind reported 57.4 per cent accuracy at classifying how far through a task a robot is, is a low number for something that sounds basic, and it is the number that decides whether a robot recovers or carries on pouring.

Speed. Correct and slow is the normal state of a demonstration.

Why this is the most important thing in robotics right now

Hardware has not been the constraint for years. You can buy a capable arm, and a competent humanoid now costs less than a car. What has been missing is a way to make a machine do a job nobody wrote down for it.

If foundation models deliver on transfer, the bottleneck moves to hands and data, and robotics starts to look like software: a general capability that improves for everybody at once, rather than an integration project rebuilt per site. That has not happened yet. It is the bet the entire field is currently placing.