Google DeepMind teaches robots to act with RT-2

Teaching a robot to clean up your desk used to be a technical nightmare.

Historically, you had to train it on hundreds of individual steps: how to recognize an apple, how to tell it apart from a red ball, and how to position its gripper to lift it. If you introduced a novel object, the robot would simply freeze.

Google DeepMind just introduced RT-2 (Robotics Transformer 2), a new vision-language-action model designed to fix that exact limitation.

Ending the robotic game of telephone

Until now, robots ran on a messy stack of disconnected systems. A vision model would inspect the room, a large language model would plan the steps, and a low-level controller would try to translate those thoughts into physical motor movements.

RT-2 collapses that entire pipeline into a single Transformer model. It takes text and images from the web and translates them directly into robotic actions.

Instead of playing an imperfect game of telephone between brain and limbs, one model handles both the high-level reasoning and the physical output.

Common sense from the web

Collecting physical training data across billions of real-world scenarios is prohibitively expensive. DeepMind's workaround is simple: borrow knowledge already sitting on the internet.

Because RT-2 is trained on massive web corpora, it already understands abstract human concepts. It knows that a crumpled chip bag or a banana peel becomes trash after you eat, even if it was never explicitly trained on grasping that specific garbage.

Across more than 6,000 robotic trials, RT-2 matched the previous RT-1 model on familiar tasks.

More impressively, in completely novel, unseen scenarios, RT-2 nearly doubled the success rate, jumping from RT-1's 32% to 62%.

Why it matters

For robotics researchers and developers, RT-2 proves that foundation models can do much more than chat—they can physically navigate the real world. By tapping into web-scale data, teams no longer need to manually record millions of physical interactions just to teach a robot basic intuition.

Across the industry, this changes the playbook for embodied AI. The traditional split between high-level reasoning software and low-level motion controllers is starting to look obsolete. DeepMind's single-model approach will likely push competitors to ditch modular stacks in favor of end-to-end architectures.

We think this is a clever architectural shortcut that solves the physical data scarcity bottleneck. However, a 62% success rate in novel environments—while a huge leap over 32%—is still far from production-ready. In real-world households or factory floors, failing nearly four out of ten times is still unacceptable, but the trajectory is undeniably clear.