Gemini Robotics 2 Brings Whole-Body Control to Apollo 2
Google DeepMind’s new VLA extends manipulation into walking and full-body humanoid control, while an on-device model targets faster adaptation across embodiments.
Google DeepMind has launched Gemini Robotics 2, extending its robotics stack from upper-body tabletop tasks to whole-body humanoid control. The new vision-language-action model controls Apptronik’s Apollo 2 from feet to fingertips, combining walking, crouching, reaching and manipulation.
Earlier Gemini Robotics releases concentrated on manipulation, cross-embodiment transfer and local adaptation. Gemini Robotics 2 moves the boundary outward: locomotion is no longer treated as a separate controller that merely carries the arms into position. Walking, balance, posture and manipulation are brought into the same action policy.
The same checkpoint was tested across Apollo 2 with two hand systems and a Franka Duo. Google’s release says it can move between whole-body tasks, multifinger dexterity and conventional gripper work without changing the core model.
The update adds Gemini Robotics ER 2 for multi-step reasoning, self-correction and multi-robot collaboration. Gemini Robotics On-Device 2 is designed to run locally and adapt to new embodiments in a few hours, typically with fewer than 200 examples.
Performance remains uneven. Google reports a 32% success rate for a dustpan task and 92% for unscrewing a bulb, while also identifying movement speed as a continuing limitation. Reporting the weaker tasks provides a clearer picture of the model’s current limits.
Apptronik’s Robot Park provides the wider data loop. Fleets of Apollo 2 robots collect real-world task data at the Austin facility, at Google DeepMind and at customer sites including Mercedes-Benz and GXO.
The model and the robot are also becoming more tightly coupled. Apptronik has raised more than $935 million to commercialise Apollo, while Robot Park is designed to produce the real-world data needed to improve Gemini. Google supplies the model, Apptronik supplies the embodiment and operating data.
Have a robotics update Korthos should review? Send news, deployments, product releases, funding rounds, research, or media to tips@korthos.xyz or reach out on X at @agkorthos.