Google DeepMind released Gemini 4 Argon for robotics. It leads the ERQA benchmark with an 87% score. The model enables zero-shot control of robots like Boston Dynamics Spot. It can create digital twins from real robot or human video feeds. Argon also turns hand-drawn sketches into simulated robot tasks. It supports closed-loop steering of video world models. The model provides multi-level hindsight labeling for dexterous manipulation. It can synthesize whole-body throwing trajectories for humanoid robots. Argon verifies physical contact forces in simulations. It exports Blender renders of simulated scenes.
What it unlocks
Physical reasoning, zero-shot robot control, Real2Sim scene reconstruction, and agentic workflows for robotics tasks.
- 87% on ERQA benchmark
- 76-DoF humanoid with 22-DoF hands
- 500 Hz MuJoCo simulation
- 136-second origami folding task
What you need to act on it
- access to Google Cloud or DeepMind APIs
Sources