How robots learn: A brief, contemporary history

The real buried lede is timing: this retrospective lands at the precise moment that gap is being actively contested by multiple labs simultaneously, making the history less academic and more diagnostic.

Read alongside Physical Intelligence's π0.7 announcement from April 16, which claims a robot brain that can perform tasks it was never explicitly trained on, and the MIT Technology Review piece stops being a history lesson and starts reading like a baseline measurement. Google DeepMind's Gemini Robotics-ER 1.6 release from April 13 adds a second data point: two serious efforts, within days of each other, both claiming progress on the generalization problem that has historically kept robots narrow and industrial. The arXiv paper on LLM generalization from April 16 is worth holding alongside both announcements, because it found that models transfer well spatially but collapse on longer planning horizons, which is precisely the kind of failure mode that has kept robots from moving beyond scripted tasks for decades.

If Physical Intelligence's π0.7 can demonstrate reliable task generalization on manipulation benchmarks that neither it nor Gemini Robotics-ER was trained against, that would be the first credible evidence the historical pattern this article describes is actually breaking. If both systems revert to narrow performance outside curated demos, the history holds.