Dexterous manipulation is the bottleneck for general-purpose robots

Google DeepMind's Gemini Robotics 2 introduces an embodied reasoning model that tackles the hardest remaining problem: dexterous manipulation. While locomotion and basic navigation are nearly solved, tasks like tying a trash bag or unscrewing a sphere-shaped bulb require coordinating over twenty joints with precise force and contact management. The team admits data is the bottleneck—teleoperation

data is costly and scarce, and human video lacks the action labels needed for robot training. Crucially, the team reveals that adding physical action tokens to Gemini's pre-trained model degrades its language and vision capabilities because physical datasets are orders of magnitude smaller than internet-scale digital data. This explains why frontier models don't directly control

robots yet. The model does show cross-embodiment transfer, meaning training on one robot type (gripper) can benefit another (dexterous hand), but the fundamental unsolved problem is acquiring millions of high-quality physical interaction trajectories. The team's deployment strategy is pragmatic: industrial environments first, where safety is controlled and tasks are semi-structured, then retail, and finally homes—likely

five to ten years out. They warn that 90% success with 10% failure (breaking a wine glass) may not be acceptable for consumers. The embodied reasoning model is available via AI Studio and an API, while the action models remain with deep partners for now. The key takeaway: the physical AGI path is fundamentally different