MOUNTAIN VIEW, CALIFORNIA — Apptronik’s Apollo 2 bends towards a watering can, closes one hand around its handle and carries it across a staged room to a low shelf. The task looks ordinary until the robot’s legs become part of the problem: walking, crouching, balancing and placing must work as one connected action.Google DeepMind introduced Gemini Robotics 2 on 30 July 2026, extending its robotics models from upper-body and tabletop manipulation to whole-body control. The company demonstrated the system on the Apptronik Apollo 2 humanoid, using both Inspire hands and five-fingered SharpaWave hands, as well as a dual-arm Franka platform.
From Feet to Fingertips
Gemini Robotics 2 is a vision-language-action model that turns visual information and language instructions into motor commands. It can control a humanoid from its feet to its fingertips, rather than treating locomotion and manipulation as separate systems.
Gemini Robotics ER 2 operates above it as an embodied-reasoning model. It interprets instructions, plans multi-step work, monitors progress and hands motor execution to a lower-level control model. A third version, Gemini Robotics On-Device 2, is designed to run locally when an internet connection or cloud latency would be unsuitable.
DeepMind says the on-device model can be adapted to a new two-armed robot with fewer than 200 examples collected over several hours. The company has shown that process on several embodiments, but it has not established that the same adaptation speed will apply to every humanoid, sensor package or workplace.
The Published Results Show the Limits
The demonstrations include Apollo 2 walking to a table, collecting a watering can and placing it in a specified bin. Other clips show the robot sealing a zip-lock bag, tying a rubbish bag and handling a light bulb with a 22-degree-of-freedom SharpaWave hand.
The underlying figures are more revealing than the edited videos. Apollo 2 unscrewed a light bulb successfully in 92 per cent of trials, but screwed one in only 36 per cent of the time. Tying a rubbish bag reached 44 per cent, sealing the zip-lock bag 40 per cent and using a dustpan 32 per cent.
Whole-body manipulation also varied by task. With Inspire hands, Apollo 2 achieved 68.4 per cent success when picking objects from a table, 76.3 per cent from a shelf and 45.7 per cent from the floor. DeepMind acknowledged that movement speed still needs improvement. These are meaningful research results, but they do not describe a robot ready for an unsupervised warehouse shift.
WIRED reported that the models were trained through a mixture of human teleoperation, video examples and simulation. The robots acted autonomously during the released demonstrations, but the capabilities should not be confused with a machine learning every task from a verbal instruction alone.
What Gemini Robotics 2 Means for Robot Buyers
For organisations evaluating humanoids, the announcement changes the questions that matter. Payload, runtime and mechanical reliability remain essential, but buyers must also examine the intelligence layer: which models are included, where they run, how they are adapted to a task and what happens when execution fails.
Apollo 2 is one of the clearest hardware beneficiaries of DeepMind’s work. Its role in the demonstrations strengthens the connection between Apptronik’s modular platform and Google’s physical-AI stack. It does not make Gemini Robotics 2 a standard Apollo feature, however. Anton Robots’ independent Apollo 2 review identifies the machine as a pilot-stage and data-collection platform rather than a conventional product with public pricing and general delivery.
A prospective customer should ask whether a proposed pilot includes model access, task-specific training, safety integration, software costs and performance data gathered in the intended workplace. The comparison is also commercial, not merely technical: Apollo 2 and Agility Robotics Digit represent different levels of deployment evidence and buyer involvement.
Availability remains narrower than the headline capability. Gemini Robotics ER 2 is available to developers through the Gemini API and Google AI Studio, while the whole-body VLA and On-Device models are restricted to early-access partners.
DeepMind says ER 2 can detect nearby people, trigger safety tools and stop a robot when someone enters its working area. Those are company-run evaluations, not proof of safe, repeatable operation in an uncontrolled workplace.
Gemini Robotics 2 advances a stubbornly difficult part of humanoid engineering: combining perception, planning, balance and dexterity without reducing the robot to isolated tricks. Its uneven success rates, limited access and slow movement keep it in the research-and-pilot category. Even so, humanoid competition is shifting beyond mechanics. The stronger platform may be the one whose hardware, models, safety systems and deployment support can work together reliably.