DeepMind introduces Gemini Robotics 2 for full body robot control
DeepMind’s Gemini Robotics 2 is a vision-language-action model for full-humanoid control, shipping as a distinct product from Gemini Robotics ER 2.
On July 30, 2026, DeepMind published Gemini Robotics 2: a vision-language-action model that turns vision and language into motor commands for full-body humanoids and multi-robot work. ByteForward already covered Gemini Robotics ER 2. This card is the whole-body VLA sibling, plus the on-device variant.
Carolina Parada’s post is the primary. Three models, one family.
VLA vs ER 2 planner roles
Gemini Robotics 2 is the VLA: pixels and words in, motor control out, feet to fingertips, two-arm robots included. Gemini Robotics ER 2 is the embodied-reasoning VLM: talk to humans, parse the room, plan multi-step tasks that last several minutes, and now coordinate a team. ER 2 is on Google AI Studio and in private preview on Gemini Enterprise Agent Platform. The VLA and on-device weights are early-access partner SKUs.
If you already filed the ER 2 planner card, do not paste it here. The new sentence is whole-body control plus collaboration, not “Gemini can see.”
Whole-body and multi-robot claims
Previous Gemini Robotics stacked upper-body tabletop work. Robotics 2 walks. DeepMind’s example: Apptronik Apollo 2 told to put a watering can in a green bin on the bottom shelf — walk, pick, walk, place. Same checkpoint on Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and a Franka Duo with a Robotiq gripper. Dexterity examples: knot-tying and ziplock seals on a 22-DoF SharpaWave; tight packing on a parallel gripper.
ER 2 adds multi-robot collaboration: different bodies talking through a workflow one robot cannot finish. DeepMind is honest that multi-finger dexterity is still the weak bar.
On-device vs partner access
Gemini Robotics On-Device 2 is the efficient VLA for no-network cells. DeepMind says it adapts to new bi-arm bodies in a few hours, typically under 200 examples — Dexmate, SO101, Trossen in the post. That is the motion-transfer claim. It is not a public download.
What early-access hardware looks like
Named partners in the acknowledgements include Apptronik, Boston Dynamics, and Agile Robots. Hardware in the demos is Apollo 2 and Franka. ASIMOV-Agentic is the new safety bench: refuse unsafe VLA tool calls, predict impossible tasks, ask a human when unsure. ER 2 is described as stronger on human-proximity stops.
Gemini Robotics 2 is a July 30 whole-body VLA. Try ER 2 in AI Studio if you want the planner. Apply for trusted tester if you want motors. Do not confuse the blog’s walking can with a product you can pip install. The safety technical report is the other primary if you are buying a cell that shares a floor with people.
