No. A robot that can coordinate its feet, torso, arms and fingers has achieved a difficult form of whole-body control, but that does not by itself mean it has a body sense like ours. Control is the ability to make many parts act together toward a goal. Body sense is the continuous perception of where the body is, how much effort it is using, what is touching it and how its internal condition is changing. The first can become impressive while the second remains narrow, engineered and unlike lived human awareness.

What changed this week

On 30 July 2026, Google DeepMind introduced Gemini Robotics 2. The system combines a vision-language-action model for motor control with a higher-level reasoning model for planning. In the company's demonstration, a humanoid can walk toward an object, crouch, pick it up, carry it and place it on a low shelf. Other demonstrations include delicate hand tasks and two robots dividing a job.

This is a real integration problem. Walking changes balance. Crouching shifts the centre of mass. Reaching changes the forces the legs must counter. Grasping requires the hand to close without crushing or dropping the object. A controller that treats these as separate tricks will fail as soon as they overlap. Whole-body control means coordinating them as one changing mechanical problem.

But coordination is not the same as sensation.

The three layers hidden inside “knowing the body”

Humans use several kinds of bodily information at once. Three distinctions help clarify what a robot demonstration does and does not establish.

External perception tells us about the surroundings: the shelf is low, the floor is uneven, another person is approaching. Cameras and other outward-facing sensors provide a robot with an engineered version of this information.

Proprioception tells us about position, movement, force and effort within the musculoskeletal system. It is why you can touch your nose with your eyes closed. Human proprioception is not a single sensor; signals from muscles, tendons, joints and skin are combined with motor commands and prior experience. Robots also use the word proprioception for internal measurements such as joint angle, velocity and motor torque. The shared word is useful, but the systems are not equivalent.

Interoception concerns the body's physiological condition: breathing, heartbeat, temperature, pain, fatigue, hunger and the broader feeling of how the organism is doing. It contributes to emotion and self-awareness in humans. A battery gauge or temperature sensor gives a robot internal state data, but reading a number is not evidence of a felt condition.

Whole-body control can use the first two kinds of data operationally without possessing the third in any human-like sense. It can also use internal measurements without cultivating the kind of perceptual discrimination a somatic practitioner develops.

A concrete example: picking up from the floor

Suppose a humanoid is told to pick up a watering can from the floor and place it on a shelf.

The system must locate the can, plan a route, place its feet, lower the pelvis, keep balance, shape the hand, estimate contact, lift, turn and reach. At every moment it needs feedback: Did the foot land where expected? Did the fingers make contact? Is the load pulling the arm off course? Whole-body control closes these loops quickly enough for the sequence to work.

A person doing the same task also feels effort accumulating in the thighs, pressure shifting across the feet, the handle biting into the fingers and the breath changing with the lift. Those sensations may alter the action before any visible error occurs. Someone with movement training may notice even finer distinctions: whether the reach begins from the shoulder, spine or support leg; whether force is distributed or locally held; whether the breath assists or interrupts the phrase.

The robot may have measurements related to some of these events. That does not mean the events are organized into the same perceptual world. The difference is not mystical. It is a difference in sensors, integration, learning history, biological need and the purposes for which distinctions become meaningful.

Why the new data matters

A second release from 30 July makes the gap visible. ACE-Data-0 records human household activity with synchronized first- and third-person video, full-body and hand motion, object trajectories, sound and tactile signals. Its authors built 75,000 interaction episodes and found that current methods still struggle with contact, occlusion, egocentric movement and long tasks.

This is progress because it stops pretending a pose sequence is the whole action. Touch, objects, viewpoints and time matter. It supports the broader turn toward systems that can accept different bodies and sensor arrangements, described in the Lab's explainer on why most movement models only know one body.

Yet even rich exterior data does not automatically produce perception. As the Lab's recent analysis argues, better sensors do not by themselves create a better perceiver. A system learns to notice the distinctions rewarded by its objectives. If training rewards task completion, it may become excellent at getting the can onto the shelf while remaining indifferent to effort quality, comfort or the organisation of the movement.

What we should say precisely

“Whole-body intelligence” is a useful engineering shorthand if it means integrated perception, planning and control across an entire robot. It becomes misleading when it quietly imports stronger claims about feeling, self-awareness or human-like embodiment.

The careful conclusion is more interesting than either hype or dismissal. Robots are gaining broader, more adaptive control of physical bodies. They are also receiving richer streams of bodily and environmental data. Those are necessary steps toward more capable embodied systems. They are not proof that a machine experiences its body, and they do not make somatic perception a solved problem.

The achievement is coordination. The open question is what, if anything, turns coordinated signals into a body that matters to itself.

References

Cao, Y., Xie, H., Wen, B., Yao, R., Liu, Y., Huang, Y., Liao, Z., Wang, Y., Liu, H., Tian, X., Su, D., Zhuo, L., Tao, D., Wang, X., Pan, L., & Liu, Z. (2026). ACE-Data-0: Human-centric ambient capture as embodied data engine. arXiv:2607.28625. https://arxiv.org/abs/2607.28625

Craig, A. D. (2002). How do you feel? Interoception: The sense of the physiological condition of the body. Nature Reviews Neuroscience, 3(8), 655–666. https://doi.org/10.1038/nrn894

Laschi, C. (2025). The multifaceted approach to embodied intelligence in robotics. Science Robotics, 10(102), eadx2731. https://doi.org/10.1126/scirobotics.adx2731

Parada, C. (2026, July 30). Gemini Robotics 2 brings whole body intelligence to robots. Google DeepMind. https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/

Proske, U., & Gandevia, S. C. (2012). The proprioceptive senses: Their roles in signaling body shape, body position and movement, and muscle force. Physiological Reviews, 92(4), 1651–1697. https://doi.org/10.1152/physrev.00048.2011