Google DeepMind unveiled Gemini Robotics 2, an updated vision-language-action model that, DeepMind says, adds intelligent whole-body control, higher dexterity across end effectors, and multi-robot collaboration. The company demonstrated the stack on Apptronik’s Apollo humanoid — referred to as Apollo 2 in DeepMind’s materials — according to reporting by The Robot Report.
Earlier Gemini Robotics demos focused on upper-body, tabletop work. The Robot Report’s summary of the new release says the model can drive walking, crouching, reaching, and object manipulation in one loop. In one Apollo example, an instruction such as “put the watering can into the green bin in the bottom shelf” leads the robot to walk to the table, pick up the can, move to the shelves, and place it — still slow by DeepMind’s own admission, but framed as whole-body coordination rather than arm-only scripting.
Alongside the main VLA, DeepMind released Gemini Robotics ER 2, an embodied-reasoning vision-language agent for multi-minute task planning and human communication, and Gemini Robotics On-Device 2, a local VLA aimed at low-connectivity sites. DeepMind claims On-Device 2 can adapt to new bi-arm embodiments in a few hours with fewer than about 200 examples; treat that as a vendor claim until independent benchmarks appear. ER 2 is listed on Google AI Studio and in a private Gemini Enterprise preview; the VLA and On-Device models are early-access partner releases.
Dexterity demos include a five-fingered, 22-DoF SharpaWave hand on Apollo for fine actions such as knot-tying or sealing a bag, and a Franka Duo two-finger gripper setup for tight packing. RobotVersus already lists Apptronik’s Apollo as Shipping and Franka Robotics as a manufacturer — this story is a software-stack announcement, not a change to Apollo’s catalogue specs or RV Score.
On safety, DeepMind introduced ASIMOV-Agentic, a benchmark for agentic refusal of unsafe tool calls and uncertainty hand-off to humans, and said ER 2 improves proximity stopping when people approach. Read The Robot Report’s write-up and DeepMind’s own materials for early-access paths; product marketing language (“unlocks”, “step change”) should stay attributed to the source.