Google DeepMind has released Gemini Robotics 2, a foundational AI model designed to give humanoid robots what the company calls “whole body intelligence.” The development marks a significant step in robotic control, moving beyond the manipulation of objects with arms and hands to enabling machines that can navigate, balance, and interact with their environment using their entire physical form. This capability is critical for developing general-purpose robots that can operate in human-centric spaces.
The new model allows a robot to process video from its own camera and a high-level text command, such as “pick up the baseball glove,” and translate it into a continuous stream of low-level motor commands for every joint in its body. According to a post on the company’s blog, this integrated approach is essential for tasks that require a robot to maintain balance while reaching, carrying objects while walking, or dynamically adjusting its posture to interact with the world.
From the Torso Up to the Whole Machine
The previous iteration of Google’s robotics model focused primarily on controlling a humanoid robot’s upper body. While this allowed for impressive feats of object manipulation, it treated the robot’s legs and torso as a static or separately controlled platform. This limitation is akin to a person trying to work while sitting in a fixed chair—functional for some tasks, but severely restrictive for many others.
Gemini Robotics 2 removes this restriction. Whole body intelligence means the AI model considers the robot as a single, interconnected system. To perform a task, it must simultaneously calculate how to move its arms, shift its weight, adjust its torso, and position its legs to remain stable and efficient. This is a far more complex computational problem, mirroring the subconscious coordination humans use for everyday activities.
In a demonstration, Google DeepMind showcased the model controlling an Apollo 2 robot from hardware partner Apptronik. The task was simple: take a baseball glove off a shelf. As reported by The Verge, executing this action required the Apollo robot to coordinate its entire body. It had to position its feet for a stable base, bend at the waist and knees to lower its center of gravity, extend its arms to reach the glove, and continuously make micro-adjustments to keep from falling over. Every part of the robot worked in concert, governed by a single AI model.
The Technical Architecture: A Closed-Loop System
The core of Gemini Robotics 2 is a closed-loop control system. This means the model is not just executing a pre-planned sequence of movements. Instead, it operates in a continuous feedback loop, making it reactive and adaptable.
The process works as follows:
- Input: The model receives a video stream from the robot’s onboard camera and a text-based instruction for the task at hand.
- Processing: The Gemini model processes this visual and language data to understand the state of the environment and the goal.
- Output: It generates real-time motor commands, specifically joint angle targets, for the entire robot. These commands are sent at a high frequency, allowing for fluid and responsive motion.
This cycle repeats continuously, enabling the robot to adjust its actions based on what it sees. If an object is moved slightly or if the robot’s footing is less stable than anticipated, the model can correct its posture and trajectory on the fly. According to Google DeepMind, this entire process runs on the robot's onboard computer, which is essential for autonomous operation without reliance on a constant, high-bandwidth connection to external servers.
This real-time, onboard processing is a key differentiator. It moves the robot from being a machine that follows a script to one that can perceive and react, a fundamental requirement for operating in unpredictable, real-world environments.
Training a Digital Brain for a Physical Body
Developing a model capable of such complex control requires a sophisticated training strategy. A purely physical approach—training a robot through trial and error in the real world—is slow, expensive, and potentially dangerous to the hardware. Google DeepMind’s solution is a hybrid approach that leverages both simulation and real-world data.
The training dataset is built from two primary sources:
Simulated Data: The vast majority of the training occurs in a physics-based simulation. Here, a digital twin of the robot can attempt a task millions of times without any physical wear and tear. This allows the model to explore a wide range of scenarios, learn the fundamentals of balance and motion, and master skills in a compressed timeframe. DeepMind states that a new skill can often be learned in simulation within a single day.
Human Demonstrations: To ground the model in reality, the simulated data is supplemented with a smaller dataset of real-world demonstrations. A human operator teleoperates the physical robot to perform a task correctly. These demonstrations provide the model with high-quality examples of how to execute a skill efficiently and safely in the physical world.
This combined dataset is used to train the model with a technique called Reinforcement Learning from Demonstrations (RLfD). The model first learns to imitate the human demonstrations (learning from demonstration). Then, it uses reinforcement learning to refine its performance, discovering new and potentially more efficient ways to achieve the goal through virtual trial and error. This method helps bridge the notorious “sim-to-real” gap, where behaviors learned in a perfect simulation often fail to transfer to the messy and unpredictable physics of the real world.
Hardware, Partners, and the Humanoid Race
While Google DeepMind is focused on the AI, the software needs a physical body to inhabit. The company is taking a hardware-agnostic approach, designing Gemini Robotics 2 to be adaptable to various humanoid platforms. The most prominent partnership highlighted in this release is with Apptronik, a robotics company that develops the Apollo humanoid robot.
This collaboration is indicative of a broader trend in the industry. The development of capable humanoid hardware and the development of intelligent AI “brains” are progressing in parallel, often at different companies. The race is on to create the definitive software stack for these machines. OpenAI has partnered with Figure AI to power its Figure 01 robot, while Tesla is pursuing a vertically integrated approach with its Optimus robot and its own internal AI development.
The emergence of foundational models like Gemini Robotics 2 suggests a future where robotics hardware manufacturers could license an AI control system, much like smartphone makers license a mobile operating system. Success will depend on how well the AI can generalize across different body types, sensor suites, and motor capabilities.
What to Watch Next
The demonstration of whole body intelligence is a significant milestone, but the work is far from complete. The tasks shown, while complex from a robotics perspective, were performed in a controlled and uncluttered lab environment. The true test for Gemini Robotics 2 and its competitors will be performance in unstructured, dynamic settings like a busy warehouse, a cluttered factory floor, or eventually, a home.
Moving forward, the key area to watch is generalization. The next objective will be to expand the model’s library of skills and, more importantly, its ability to combine those skills to solve novel problems it has not been explicitly trained for. Can a robot trained to pick up a glove and open a door synthesize those abilities to open a cabinet and retrieve a tool? This capacity to reason and adapt is the ultimate goal of general-purpose robotics.
The industry will also be looking for evidence of economic viability. The long-term promise of humanoid robots is to supplement human labor in tasks that are dangerous, repetitive, or physically demanding. The next phase of development will require moving from impressive tech demos to reliable, real-world applications that provide tangible value. The progress of Gemini Robotics 2 will be measured not just by the complexity of the tasks it can perform, but by its reliability and utility outside the lab.