Robotics and Embodied AI

Robotics is uniquely hard for AI because: data is scarce (no internet of robot demonstrations), the real world is continuous and high-dimensional (vs. discrete tokens), simulation-to-reality transfer is imperfect, and physical actions have irreversible consequences.

LLM integration has improved generalization dramatically. Models like RT-2 and Google's PaLM-E use pre-trained vision-language models as the "brain" of the robot, enabling generalization to novel objects and instructions without task-specific training.

Humanoid robots (Figure, 1X, Unitree) are receiving significant investment because human-form robots can operate in human-designed environments and use human-collected demonstration data more directly.