Google's Gemini 4 Argon scores 87% on ERQA and controls robots in simulation and on hardware
- Google's Gemini 4 Argon scores 87% on the ERQA embodied reasoning benchmark using pure reasoning, which the post says beats every other frontier model.
- Argon drives a Boston Dynamics Spot quadruped with a 6-DoF arm to learn a foot-pedal trash bin task, first practicing in a MuJoCo digital twin of the bin before the physical attempt.
- For Real2Sim, Argon turns an egocentric human video into a simulation of a 76-DoF bimanual humanoid with 22-DoF five-fingered hands, running at 500 Hz in MuJoCo contact physics.
- In zero-shot control, Argon reasons over a visual coordinate harness and action primitives to upright a fallen mug and put a napkin inside over several turns on a simple robotic arm.
- In a closed loop, Argon rewrites timestamped action prompts from rollout keyframes to steer a video world model into spelling A-R-G-O-N with Scrabble tiles over a 33-step rollout.
Hacker News 의견들
Aren't we still like a decade away from running this on battery on a robot that isn't the size of a school bus? Maybe 5-7 years for a plug-in robot, tops.
Honestly the planning doesn't need to live on the robot. Local models handle locomotion and balance, and the big model phones home for planning. Latency is fine for most jobs.
Partly disagree. Robot subsystems are time critical, and a couple of kW for the model is the real problem. You want the LLM high level and the subsystems autonomous.
The zero-shot visual harness approach is the part I found interesting. Plenty of tasks don't need a fancy 6DoF dexterous arm, just a generalist model calling predefined actions.
I have a Unitree humanoid and Argon handled obstacle avoidance way better than Opus or Astra, which kept crashing into tables. Pretty scary for something I vibecoded at home.
Google should be wiping the floor with the competition with all those resources. Figure 3 looks better, and Unitree robots are doing parkour while this one hangs from a hook.