Google Gemini Robotics 2 gives humanoid robots full-body AI control
The company demonstrated the technology on humanoid and two-arm robots, including Apptronik’s Apollo 2. In one example, the humanoid followed an instruction to walk towards a table, pick up a watering can and place it on the bottom shelf of a storage unit. Google accepts that movement speed still needs work, which means these robots are not ready to walk into every home or factory just yet.
Three models handle movement, planning and local tasks
The Gemini Robotics 2 family has three main models.
Gemini Robotics 2 is a vision-language-action model. It converts camera input and spoken or written commands into physical movements. It can control a humanoid from its feet to its fingers and operate two-arm robots.
Gemini Robotics ER 2 works as the high-level brain. Google says it can speak with people, examine a room, plan steps and monitor a task until completion.
Gemini Robotics On-Device 2 runs locally on robot hardware. It can be adapted to a new two-arm robot using fewer than 200 examples and a few hours of training data.
Gemini Robotics ER 2 is available through the Gemini API and Google AI Studio. It is in private preview on the Gemini Enterprise Agent Platform. The movement and on-device models are limited to early-access partners.
Robots can use hands for delicate work
Google says Gemini Robotics 2 can control the five-fingered SharpaWave hand used with Apollo 2. The hand has 22 degrees of movement and was tested on jobs such as tying knots and closing ziplock bags.
The model can operate standard two-finger grippers too. On a Franka Duo setup, robots were shown packing objects into tight spaces. These tasks look simple to people, yet they demand careful pressure, timing and finger placement from a machine.
Gemini Robotics ER 2 tracks progress in real time
A robot must know whether it finished a job properly. Gemini Robotics ER 2 watches continuous video and divides task progress into five stages.
Google reported 57.4 per cent accuracy in progress classification tests. The model recorded 91.3 per cent accuracy in identifying the exact moment an event happened, with an average error of 0.96 seconds. Google claims this process runs four times faster than larger model categories tested by the company.
The model can plan its next action as the robot continues moving. This removes some of the awkward pauses seen when a machine stops to process each new step.
Multiple robots can now work together
Gemini Robotics 2 introduces communication between different types of robots. A wheeled machine could inspect an indoor area, then pass a task to a humanoid or robotic arm that is better suited to lifting and handling objects.
Google demonstrated this approach using Apollo 2 and the Franka F3 Duo. The shared AI system divided the work and tracked the handover between machines.
Safety remains a major test. Google has introduced the ASIMOV-Agentic benchmark to measure whether a robot refuses unsafe instructions, recognises an impossible task and asks a person for help. Gemini Robotics ER 2 can detect nearby people and stop a humanoid when someone moves too close.