Most striking AI progress has been in language and images. But much of intelligence involves acting in the physical world.
The Embodiment Argument
Some researchers argue that true general intelligence requires interacting with a physical environment — learning cause and effect, physics and common sense through experience. Others think much can be learned from data alone.
Robotics Progress
Robots are benefiting from the same methods as language models: large models trained on diverse data, vision-language-action models linking perception to movement, and learning in simulation before real-world deployment.
Challenges
- Scarce real-world robot data compared with text.
- Physical safety and reliability requirements.
- The difficulty of dexterous manipulation in varied environments.
- Hardware cost.
World Models
Systems that learn to predict how environments change — used for planning, simulation and video generation — are seen by some as a key ingredient for general intelligence.
Implications
A system excelling at cognitive tasks but unable to act physically would transform knowledge work while leaving much manual work untouched. Progress in robotics will shape AI's broader economic impact.