Google DeepMind has taken another step toward making humanoid robots feel less like carefully choreographed machines and more like general-purpose helpers. In a new update to its Gemini Robotics line, the company says Gemini Robotics 2 can now control “entire humanoid robots,” expanding beyond the earlier focus on upper-body movement. The practical implication is straightforward but significant: if an AI system can coordinate motion from feet to fingertips, it can handle tasks that require posture changes, balance management, and precise hand placement at the same time—exactly the kind of complexity that makes real-world robotics so difficult.
The announcement centers on what DeepMind describes as whole-body motions. Where previous iterations were geared toward controlling parts of a robot’s body—often the arms, hands, and related upper-body degrees of freedom—Gemini Robotics 2 is designed to support actions that involve the full physical stack. That includes walking, crouching, stretching, and manipulating objects while the robot’s legs and torso adjust to maintain stability and reach. In other words, the model isn’t just deciding how to move a hand; it’s also deciding how to get the body into the right configuration to make that hand movement possible.
This shift matters because everyday tasks rarely fit neatly into “arm-only” scenarios. Consider something as mundane as picking up an item from a shelf. A robot can’t simply extend an arm and hope for the best. It must approach the shelf correctly, align its torso, bend or crouch to reduce awkward reach, and then execute a grasp with enough precision to avoid dropping the object or colliding with nearby items. If the robot’s legs and hips don’t move in coordination with the arm, the grasp becomes either physically impossible or unreliable. Whole-body control is the difference between reaching and actually performing.
DeepMind’s examples, shared alongside the update, illustrate this broader capability using Apptronik’s Apollo 2 robot. In one demonstration, the robot bends over to pick up a watering can. In another, it finds and takes specific items off a shelf. These are not just “cool clips.” They represent a class of behaviors that require multiple coordinated subsystems: locomotion or stance adjustment, perception of where objects are relative to the robot, and then fine motor control for grasping. The key claim is that Gemini Robotics 2 can tie these elements together through whole-body intelligence rather than treating them as separate stages.
To understand why this is a big deal, it helps to look at what robotics engineers typically do today. Many systems rely on pipelines: perception estimates object location, planning computes a path and target pose, and then control executes joint trajectories. Even when the pipeline is efficient, it can be brittle. If the environment changes slightly—an object is moved, the shelf height differs, the robot’s starting position is off by a few centimeters—the system may fail or require extensive re-planning. Learning-based approaches aim to reduce that brittleness by letting models generalize across variations. But learning-based control still struggles when the action space becomes too large.
Whole-body control expands the action space dramatically. A humanoid robot has many joints, and coordinating them while maintaining balance is inherently harder than controlling a subset of joints. The robot must manage center-of-mass shifts, avoid tipping, and keep its feet planted or moving safely. At the same time, it must place its hands accurately enough to manipulate objects without collisions. When DeepMind says Gemini Robotics 2 supports motions “ranging from its feet to fingertips,” it’s essentially claiming that the model can handle this coupled problem rather than only mastering isolated components.
There’s also a subtle but important point about how this changes the user experience. Upper-body-only control can still feel limited: the robot might be able to reach and grasp, but it may not be able to reposition itself effectively. Whole-body control enables a more natural interaction style. Instead of requiring the robot to be placed in a perfect starting stance, the system can adapt its posture as part of the task. That means fewer constraints on where the robot begins and fewer assumptions about the environment. In real deployments—warehouses, offices, homes—those assumptions are often wrong.
DeepMind’s framing suggests that Gemini Robotics 2 is built to support a wider range of actions precisely because it can coordinate the entire body. Walking and crouching are obvious examples, but stretching and manipulating objects are where the complexity really shows. Stretching isn’t just extending an arm; it involves torso rotation, hip alignment, and careful balance. Manipulation isn’t just closing a gripper; it involves approaching the object with the right orientation, compensating for slight misalignments, and maintaining stable contact. Whole-body control provides the physical context that makes those fine manipulations feasible.
What makes this development particularly interesting is the direction it points for the robotics field. For years, humanoid robotics has been pulled between two extremes. On one side are highly engineered systems that can perform specific tasks reliably but struggle to generalize. On the other side are learning systems that can generalize better but often require careful training setups and can be unpredictable in edge cases. Whole-body control sits at the intersection: it’s a step toward generalization, but it also demands reliability because balance and safety are non-negotiable.
If Gemini Robotics 2 truly enables consistent whole-body behavior, it could reduce the gap between “robot can do the task in a demo” and “robot can do the task repeatedly in messy conditions.” Demos tend to hide the hard parts: the exact starting position, the lighting conditions, the object placement, the friction properties, and the small mechanical tolerances that affect real grasps. Whole-body control doesn’t eliminate those issues, but it gives the robot more degrees of freedom to recover. If the robot can adjust its stance and posture dynamically, it can compensate for imperfect initial alignment. That’s a form of robustness that purely arm-based control can’t provide.
There’s also a broader strategic angle. DeepMind’s Gemini Robotics line is part of a larger trend in AI: moving from models that interpret the world to models that act in it. The more an AI system can translate high-level intent into coordinated physical behavior, the more it can serve as a general interface between human goals and robotic execution. Whole-body control is a major milestone in that translation layer. It turns “do X” into “move your body in a way that makes X physically achievable.”
This is where the unique take comes in: whole-body control isn’t only about adding more joints to the model. It’s about changing what “understanding” means for a robot. When a system controls only the upper body, it can treat the rest of the robot as a fixed platform. But when it controls the entire body, the robot must implicitly understand how its own motion affects reach, stability, and the geometry of interaction. That internal coupling forces the model to learn relationships between posture and manipulation outcomes. In effect, the robot learns a kind of embodied reasoning: not just where the object is, but how the robot’s configuration determines what it can do next.
That embodied reasoning is likely to be crucial for scaling to more complex tasks. Picking up an object is one thing. Carrying it while walking, turning around, navigating around obstacles, and then placing it precisely somewhere else are much harder. Those tasks require continuous coordination between locomotion and manipulation. Whole-body control is the foundation for that continuity. Without it, the robot would have to stop, reorient, and restart—turning fluid assistance into a sequence of brittle steps.
DeepMind’s update also highlights the role of humanoid platforms in accelerating progress. Apptronik’s Apollo 2 is used in the showcased examples, and that choice underscores a practical reality: progress in AI robotics depends not only on algorithms but also on hardware that can execute complex motions reliably. Humanoid robots are mechanically challenging, but they offer a standardized morphology that can be leveraged for learning. If a model can control a humanoid’s whole body, it can potentially transfer skills across tasks that share similar physical constraints—balance, reach, and contact-rich manipulation.
Still, it’s worth being clear about what this announcement does and does not guarantee. Whole-body control capability in a set of demonstrations is not the same as universal competence across all environments and all objects. Real-world robotics includes unpredictable factors: slippery surfaces, deformable objects, occlusions, clutter, and varying object sizes and weights. Even with whole-body intelligence, the robot will need robust perception and safe control policies. The announcement focuses on the model’s ability to generate whole-body motions, but the broader system performance will depend on how perception, planning, and low-level control integrate with the model.
Even so, the direction is unmistakable. DeepMind is moving from partial control to full-body coordination, and that is exactly the kind of incremental leap that can unlock new categories of tasks. Walking and crouching are not just extra features; they are the mechanisms that allow a robot to change its relationship to the environment. Once a robot can change that relationship, it can access objects that would otherwise be out of reach or require awkward arm angles. It can also avoid collisions by repositioning its body rather than forcing the arm to contort.
For industries watching humanoid robotics, the takeaway is that the “capability surface” is expanding. Robots that can only control the upper body are limited to scenarios where the robot is already positioned well. Robots with whole-body control can operate more autonomously in less structured settings. That autonomy is what drives cost reduction and scalability: fewer manual interventions, fewer resets, and less reliance on perfect initial conditions.
There’s also a human-centered implication. People interact with objects in ways that involve body movement—bending down, reaching across, shifting weight, stepping closer. A robot that can mirror those whole-body adjustments can feel more natural and less like a machine that’s constrained to a narrow set of poses. That matters for adoption, because users don’t want to think in terms of robot limitations. They want the robot to adapt.
DeepMind’s announcement, therefore, should be seen as both a technical update and a signal about where robotics