One big difference between LLM progress and robotics is that language models benefited from purely digital feedback loops — training, testing, and scaling could all happen in simulation.

Robotics still pays a heavy “reality tax”. Every improvement eventually has to survive messy, unstructured physical environments where sensing, actuation, and safety interact in unpredictable ways.

My guess is we’ll see a ChatGPT-like moment first in semi-structured environments (warehouses, logistics, industrial assistance) rather than homes.

A general household robot feels closer to a GPT-4 equivalent problem than a GPT-3.5 one because reliability matters much more when failure affects the physical world.