• 1950 — Turing’s ideas about testing intelligence beyond abstract puzzles (embodied Turing-test lineage cited in surveys).
• 1980s–90s — Brooks and others argue intelligence need not rely on a complete world model and should be grounded in sense–act loops.
• 2010s — Deep RL and simulators (MuJoCo, Gazebo, Habitat) enable large-scale training; navigation and manipulation datasets grow.
• 2022– — RT-1/RT-2, PaLM-E, OpenVLA bring web-scale vision–language knowledge into control; world models and humanoids accelerate in parallel.

Primarily from arXiv:2407.06886 introduction and related citations; a public timeline, not oral history.