DeepMind teaches Apollo 2 cornhole from one video — the frontier lab's Video-to-Sim-to-Real pipeline
From the Robots digest for 2026-10-14
What happened: Google DeepMind's Video-to-Sim-to-Real pipeline — running on Gemini 4 Argon, the model shipped September 30 — taught Apptronik's Apollo 2 humanoid to play cornhole from a single video, a demonstration in which the robot sank seven straight cornhole bags (Crypto Briefing, ~4:46 PM ET October 14).
Development vs debut: development of the humanoid-reality-check war's Gemini Robotics layer, which entered the record October 3 with Gemini Robotics research lead Keerthana Gopalakrishnan's "GPT-2 stage" calibration and the three-model Gemini Robotics 2 architecture — and of the Gemini 4 ship-window thread on file. The new artifact is the video-first skill-acquisition datapoint: one video in, a physical game out — demo-grade, not a deployment, and the company's own showcase. It sits in direct tension with the same record's economics floor (the Anthropic report's 0.3%-cheaper figure) as a capability-side exhibit.
Watch: sim-to-real skill acquisition beyond a retry-tolerant single-game demo; independent verification of the seven-straight run; whether Apollo 2 pairs the demo with a deployment record.
Links: Crypto Briefing