Gemini 4 Argon Turns One Video Into a MuJoCo Sim, Apollo 2 Hits 7 Straight Shots
TechPreacher · x · 2026-10-26
Google DeepMind demonstrated its Video-to-Sim-to-Real pipeline: fed one ordinary clip of a person playing cornhole, Gemini 4 Argon built the entire training loop, and Apptronik's Apollo 2 humanoid sank seven consecutive shots on real hardware.
Pipeline (no mocap, no robot data):
- Perceive: video understanding maps board, bags and thrower motion, reasons about throw physics
- Build sim: authors a MuJoCo scene of task and robot as code
- System identification: calibrates simulated dynamics to match real hardware
- Train in sim: computes whole-body joint motion and practices launches
- Transfer and refine: executes on the real robot, improves after a few real throws
DeepMind claims this compresses weeks of manual simulation modeling into hours; Kanishka Rao's developer post shows Argon can reconstruct calibrated MuJoCo twins from three sources, including robot teleop episodes. The point isn't cornhole — it's a general-purpose frontier model doing the robotics engineer's job end to end.
More from Embodied
- RoboJudge: 800 human-annotated videos benchmark MLLM judges for embodied video generation — CUHK-CSE · 2026-10-26
- HKUST transfers Unitree G1 controller to full-size H2 with one GPU in ~8 hours — FinanceYF5 · 2026-10-26
- China Mobile shows non-invasive BCI + AR glasses controlling a robot hand — FinanceYF5 · 2026-10-26
- Berkeley diffusion model generates RL policy weights, letting a real G1 dance — FinanceYF5 · 2026-10-26
- KAIST Humanoid v0.9 pushes 35 kg and lifts 20 kg with onboard vision — FinanceYF5 · 2026-10-26
- Figure F.03 logs 422 miles across 13 days — and the treadmill test continues — FinanceYF5 · 2026-10-26