Google DeepMind is moving its robotics work from partial-body control toward full humanoid coordination. The company says Gemini Robotics 2 can “control entire humanoid robots,” a shift from the previous model, which focused on a humanoid robot’s upper body.
The update matters because many useful robot tasks are not just about grabbing an object. They require posture, balance, reaching, hand movement, and awareness of the surrounding space to work together as one system.
Whole-body control expands the action set
Gemini Robotics 2 now supports “whole-body motions” from a robot’s feet to fingertips, according to an announcement on Thursday. That means the model is designed to coordinate more of the machine at once, rather than treating the arms and hands as the main area of control.
Google DeepMind says the model enables humanoid robots to walk, crouch, stretch, and manipulate objects. Those movements are basic in human terms, but they are important for robots that need to interact with shelves, floors, bins, household items, tools, or other objects placed outside a narrow working zone.
Videos shared by Google show Apptronik’s Apollo 2 robot bending over to pick up a watering can. Another example shows the robot finding and taking specific items off a shelf.
Those demonstrations point to the practical difference between upper-body movement and whole-body coordination. A robot that can reach with its arms may still struggle when the task requires lowering its torso, changing stance, or using its hands while its body is in motion.
Movement still has limits
Google DeepMind is not presenting the work as finished. The company says its robots “have more to advance in movement speed.” That caveat is important: fuller control does not automatically mean fast, smooth, or production-ready motion in every setting.
Still, the company frames Gemini Robotics 2 as progress toward more demanding robot work. Google DeepMind says the update “is an important step towards the skills needed to complete more complex, real-world tasks that require whole-body coordination.”
In plain language, the model is meant to help robots connect high-level instructions with the physical actions needed to carry them out. A task may require the robot to move through a space, shift its posture, identify the right object, grasp it correctly, and place it somewhere else. Gemini Robotics 2 is aimed at that broader chain of action.
Dexterity gets more attention
The update is not only about legs, posture, and body movement. Google DeepMind also says Gemini Robotics 2 supports better dexterity by controlling more complex, five-fingered hands.
That hand control opens the door to actions that require more than a simple clamp-like grip. The examples given by Google DeepMind include sealing a Ziploc, tying a trash bag, and unscrewing a lightbulb.
These tasks are small but revealing. They involve flexible objects, twisting motions, pressure control, and changes in grip. For a humanoid robot, that kind of dexterity can be just as important as walking or crouching, because many everyday tasks depend on hands that can adapt to different object shapes and materials.
- Gemini Robotics 2 supports walking, crouching, stretching, and object manipulation.
- It can coordinate movement from feet to fingertips.
- It adds support for more complex, five-fingered hands.
- Google’s examples include a watering can, shelf items, a Ziploc, a trash bag, and a lightbulb.
Embodied reasoning helps robots plan tasks
Google DeepMind is also updating Gemini Robotics ER, short for embodied reasoning. This vision-language model helps robots analyze their surroundings, process instructions, and perform multi-step tasks.
Gemini Robotics ER 2 is described as better at completing tasks over an extended period of time. Google DeepMind also says it “now understands when tasks begin and end.”
That capability is central to making robots useful outside tightly controlled demonstrations. A robot does not only need to recognize a command; it needs to know what the work involves, when it has made progress, and when the job is complete.
The update also supports coordination between multiple robots of different types. Google DeepMind says the system can let different robots work together to complete tasks. In one video, Apollo 2 instructs Google’s dual-arm robot to put tools inside a bin while cleaning the garage.
Safety and local operation remain part of the picture
Google DeepMind calls Gemini Robotics ER 2 its “safest robotics model to date.” The company says it can “better detect when humans are nearby, trigger safety tool calls and bring the robot to a safe stop if someone approaches too closely.”
That safety behavior is especially relevant for humanoid robots, because whole-body movement can bring a machine into closer contact with people and objects around it. The more a robot can move, crouch, reach, and manipulate items, the more important it becomes for the system to understand nearby human presence.
Google DeepMind has also improved its Gemini Robotics On-Device Model, which can run locally on a robot without an internet connection. The updated model can adapt to new embodiments faster, including robots with “drastically different shapes, sensors and degrees of freedom.”
Taken together, the announcements show Google DeepMind pushing Gemini Robotics in several directions at once: fuller body control, more capable hands, longer task execution, robot-to-robot cooperation, safer stopping behavior, and faster adaptation to different robot forms. The central idea is clear: a useful robot needs more than a smart instruction system. It needs the physical coordination to turn instructions into action.