Gemini Robotics 2: Implementing Whole-Body Intelligence in Robotics

Google DeepMind | July 30, 2026 | By Carolina Parada

For decades, the dream has been to create robots that can seamlessly integrate into human environments and provide meaningful assistance.

Historically, the field of robotics has been limited. Most machines were pre-programmed for rigid, repetitive sequences or relied on teleoperation for narrow tasks. They lacked the capacity for autonomous learning or the flexibility to handle unpredictable settings. Furthermore, the "transfer problem"—moving a learned skill from one physical chassis to another—was a massive hurdle.

To solve these challenges at scale, robots regardless of their form factor require AI that allows them to think, act, and interact safely. While the original Gemini Robotics showed how multimodal understanding could drive action, Gemini Robotics 2 represents the next evolution: an intelligence layer for truly adaptable machines.

---

🤖 The Architecture of Intelligence

Gemini Robotics 2 isn't a single model, but a suite of three specialized models designed to handle different aspects of embodiment.

Model Comparison Matrix

Detailed Model Breakdown

• Gemini Robotics 2: This is the powerhouse VLA. It translates visual and linguistic inputs directly into motor commands. It manages everything from the feet to the fingertips of a humanoid, bringing unprecedented dexterity to both multi-fingered hands and standard grippers.
• Gemini Robotics ER 2: The "brain" of the operation. As a Vision Language Model (VLM), it handles the reasoning side—understanding the physical world, planning tasks that span several minutes, and coordinating teamwork between multiple robots.
• Gemini Robotics On-Device 2: A streamlined VLA optimized for local hardware. Its standout feature is the ability to adapt to a completely new robotic embodiment using only a small amount of data collected over a few hours.

---

🏗️ System Workflow

The interaction between the reasoning layer and the action layer can be visualized as follows:

---

🚶 Humanoids in Motion: Whole-Body Coordination

Human environments are designed for human geometry, requiring the ability to bend, stretch, and balance in cramped spaces. While previous iterations focused on "table-top" tasks (upper-body only), Gemini Robotics 2 introduces whole-body control.

Example Scenario: When using the Apptronik Apollo 2 humanoid, a user might provide the following command: "Put the watering can into the green bin on the bottom shelf."

The Execution Process:

• Process linguistic intent.
• Navigate to the table.
• Grasp the watering can.
• Walk to the shelving unit.
• Precisely place the object in the target bin.

While movement speed is still an area for improvement, this represents a pivotal shift toward robots that can navigate and manipulate the real world holistically.

---

🖐️ Precision and Dexterity

Utility in the home or office requires finesse. Gemini Robotics 2 enhances dexterity across various "end effectors" (the parts that touch the world).

1. High-DOF Hands: The model can control the SharpaWave hand on the Apollo 2, which features $22$ degrees of freedom. This allows for intricate tasks such as:
   • Tying knots.
   • Sealing ziplock bags.
2. Parallel Grippers: It also maintains high proficiency with standard two-fingered grippers (e.g., on the Franka Duo platform) for complex manipulation.

---

🧠 Agentic Reasoning & Collaboration

Complex real-world goals are rarely single-step. Gemini Robotics ER 2 acts as the executive function, managing the complexity of long-term tasks.

The ER 2 Logic Loop:

while task_not_complete:
    observation = observe_environment()
    plan = reason_next_step(observation, user_goal)
    action_command = send_to_VLA(plan)
    execute(action_command)
    if failure_detected():
        self_correct(failure_type)

This allows robots to not only execute multi-step plans but also to self-correct when a step fails and generalize their behavior to entirely new goals. Furthermore, this model enables robots to function as a cohesive team to increase efficiency.

---

📊 Performance and Availability

Testing across three different embodiments (Apptronik Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and Franka Duo with Robotiq grippers) showed varying success rates:

• Whole-Body Tasks: Medium to High Success.
• Gripper-based Dexterity: Medium to High Success.
• Multi-finger Manipulation: Still challenging (Ongoing research).

Access Information

For those looking to integrate these models into their own hardware, detailed guides are available on the Developer blog.