DeepMind introduced Gemini Robotics 2 for whole-body robot control

Google DeepMind announced Gemini Robotics 2, a model aimed at whole-body control, dexterous manipulation and teamwork. The July 30 post contrasts the work with robots programmed or teleoperated for narrow task sequences and describes an effort to transfer learned capabilities across different physical embodiments. It is a separate layer from the ER 2 embodied-reasoning model announced the same day.

Embodiment is part of the model contract

A robot policy cannot be evaluated independently of mass, reach, joint limits, sensors and control frequency. Transfer between bodies is valuable precisely because those details differ, but successful transfer in selected demonstrations does not eliminate per-platform calibration. A team adopting the model needs to define which observations and actions are normalized and which remain specific to each robot.

Whole-body tasks also couple locomotion and manipulation. Reaching safely may depend on stance, balance and the motion of nearby people. Local constraints need authority to stop or reshape a model-proposed trajectory.

Build an evidence ladder for physical deployment

Begin with recorded trajectories and simulation, then progress through fixtures, restricted workspaces and supervised trials. Evaluate off-nominal loads, slippery surfaces, sensor dropout, object breakage and recovery after interruption. Track near misses and safety-controller overrides, not only completed tasks.

Version the model together with robot calibration and controller firmware so a rollback restores a tested combination. Keep human stop controls independent of the model and network path. DeepMind’s announcement marks broader ambition for adaptable physical agents; production use will be determined by the rigor of embodiment-specific testing.