DeepMind introduced Gemini Robotics 2, a system that extends beyond the previous version's focus on humanoid upper-body and tabletop tasks to control whole-body behaviors. The new model can walk, crouch and reach, handling full-body motion sequences rather than only limited upper-body actions.
Capabilities
- A single model checkpoint drove three completely different robots, demonstrating that the model is not tuned to a single physical platform but has cross-body control abilities.
- Adapting the model to a new body—one with different shapes, sensors and joints—required only a few hours and under 200 examples. This rapid adaptation improves practical deployability.
- Cross-body generalization first appeared in Gemini Robotics 1.5; Gemini Robotics 2 represents a significant advancement of that capability.
Limits and caveats
DeepMind and Google are explicit about current limitations: fine multi-finger manipulation remains challenging, and motion speed still lags behind the best human or specialized hardware solutions. In short, generalization has advanced substantially, but the full range of dexterous, fast manipulation is not yet solved.
Why this matters in practice
For decades, robotics development largely meant one body and a bespoke, hand-tuned controller per machine. Gemini Robotics 2 challenges that model by extracting intelligence from hardware and turning it into a portable "brain" that can be dropped into different bodies. The shift echoes what large language models did for software: one general model replacing many task-specific systems.
If this trend continues, strategic value in robotics may move from owning particular physical bodies to owning models that generalize across bodies and can be rapidly adapted to new hardware.
Summary
Gemini Robotics 2 may represent robotics' GPT moment: a single brain that can run multiple bodies and quickly adapt to new ones. However, important technical gaps—especially fine-fingered manipulation and speed—remain to be closed.
