Embodied AI

Core thesis

The convergence of foundation models with physical systems. Vision-language-action (VLA) foundation policies – Physical Intelligence pi0, NVIDIA GR00T, Google Gemini Robotics, Figure Helix – now translate natural-language instruction and perception into motor control, with NVIDIA Isaac providing the simulation-to-reality pipelines. The historical gap was that language models reasoned well about words but poorly about physics; closing it is the prerequisite for genuinely useful humanoid robots, autonomous drones and surgical systems. The open question has shifted from capability to deployment economics.

State of the art (2026)

Robots that take instructions in ordinary language have moved from research demonstrations to early products. The main approach is one large AI model that turns what the robot sees and is told into movement. Physical Intelligence released its model, called π0, for anyone to use. Nvidia released an open model for humanoid robots, GR00T N1. Google has Gemini Robotics and Figure has its own model, Helix. Figure raised more than $1 billion at a $39 billion valuation and is building its Figure 03 robot at its own factory. 1X opened a factory in Hayward, California, for its NEO home robot, priced at $20,000 or $499 a month. 1X says the first units went to its own staff to test at home, with deliveries to customers planned for 2026. We have not seen confirmation that customers have received them. What is not yet known is whether these robots are reliable enough, and cheap enough to run, to pay for themselves.