A look at Robotics at Google, the company's latest robotics program, which brings machine learning to robots simpler than Boston Dynamics' machines
Context & Ripple Effects
Google's 2019 robotics push lands after a long reset: back in 2015, its robotics group was spun into a separate Alphabet division after early hardware ambitions stalled. The new program's stated bet — machine learning applied to simpler robots rather than Boston Dynamics-style acrobatic machines — reads at the time like a retreat from spectacle.
The later coverage shows the bet compounding: Code as Policies in 2022 used AI models to generate robot task code, then RT-2 in 2023 trained a vision-language-action model on web text and images that could output robotic actions directly. By 2026, Alphabet folded Intrinsic, its industrial-robot software unit, into Google alongside DeepMind — confirming the center of gravity had moved from robot bodies to robot brains.
First-order effects
- Google's robotics team competes on learning algorithms rather than hardware showmanship, positioning itself against Boston Dynamics' physics-first machines instead of imitating them.
Second-order effects
- The simpler-robots strategy makes web-scale training data usable for robot control — the path that later produced RT-2's vision-language-action model and gave Google a software moat hardware rivals lack.
Third-order effects
- If the pattern holds, value in robotics consolidates around foundation-model software layers — where Google now sits with DeepMind and Intrinsic — while pure hardware specialists become platforms others program.
The trend: Robotics is shifting from a hardware-capability race toward foundation-model software stacks, with Google's simple-robots-plus-machine-learning bet as an early data point.