"The field is still GPT-2": Gemini Robotics' lead prices the hype against her own roadmap
From the Robots digest for 2026-10-03
What happened: Keerthana Gopalakrishnan, Google DeepMind's research lead for Gemini Robotics, on The Cognitive Revolution (piece in feed Oct 3): despite viral sprinting humanoids, the field is at a "GPT-2 stage" — cross-embodiment generalization and few-shot learning unsolved. Locomotion's progress is a "low-value sideshow" (sim-friendly rigid-body physics: "a lot of my friends are very productively employed… we don't run faster than Usain Bolt"); manipulation is the commercial frontier and stays bottlenecked by contact dynamics simulators model poorly. She detailed a three-model Gemini Robotics 2 architecture splitting embodied reasoning from action control (incl. On-Device 2), framed safety as a capability, and named reliability — not capability — the binding constraint (compounding error rates in multi-step tasks), countering market forecasts of useful humanoids by mid-2027. Near-term commercial reality: retry-tolerant simple manipulation, pick-and-place, close to deployment-ready. Also on record: her rebuttal to NVIDIA's Jim Fan on egocentric video, and hands overtaking grippers at the dexterity frontier.
What's genuinely new: a calibration datapoint with unusual provenance — this is the company whose own world-model bet (Genie) headlines the optimistic camp, delivering the Agility-humility-war's strongest quantitative-adjacent argument from inside a frontier lab. It directly brackets the hype layer this file tracks (XPeng IRON's year-end mass production, Tesla's ramp claims) and matches Micron's 2028 component-constraint timeline: reliability plus supply, both dated.
Links: writeup (BigGo/FIN) — podcast-adjacent source; verify quotes against the Cognitive Revolution episode before quoting at length.