Gemini Robotics 2 brings whole body intelligence to robots

For decades, we’ve dreamed of robots that can seamlessly step into our world and lend a hand. Now, that vision takes a significant stride forward.

Most robots are pre-programmed or teleoperated for narrow, repetitive task sequences. They lack the ability to truly learn for themselves or adapt to unpredictable environments. Moreover, transferring learned skills from one robot body to another remains incredibly difficult. To take on the hardest problems at scale, robots of every shape and size need AI models giving them the ability to think, act, and interact intelligently to safely complete tasks.

Introducing Gemini Robotics 2 – the intelligence layer powering the next generation of truly adaptable robots. As it takes its first literal steps, this major advance unlocks intelligent whole-body control, advanced dexterity, and multi-robot collaboration.

Learn more: https://deepmind.google/gemini-robotics

Google DeepMind demonstrates how their generalist AI model powers the Gemini Robotics 2 platform. By coordinating whole-body movement, fine motor control for dexterous tasks, and multi-robot collaboration, the team explores how these systems can navigate complex, real-world environments with greater autonomy and intelligence.

The year 2026 is widely recognized as a pivotal transition period for humanoid robotics, moving from experimental research and pilot testing toward mass deployment. Experts suggest we are at a “convergence moment” where hardware advancements and embodied AI are finally enabling robots to operate effectively in human-centric environments.

Key Trends Shaping the Future:

Intelligence and Whole-Body Control: As highlighted by advancements like Gemini Robotics 2 (1:00–1:05), the shift is moving away from pre-programmed, repetitive movements toward “whole-body intelligence.” This allows robots to reason, adapt to changing environments, and perform complex tasks—such as manipulating delicate objects or collaborating with other robots (1:30–2:28).
Transition to Industry and Home: 2026 is seeing a shift in deployment. Organizations are increasingly using humanoid robots to solve labor shortages and improve operational flexibility in manufacturing, logistics, and healthcare. Projections suggest that within coming decades, millions—potentially billions—of these robots could be integrated into human spaces, including homes.
Economic Viability: With operating costs becoming increasingly competitive (some estimates project costs as low as $2 per hour), humanoid robots are transforming from expensive research prototypes into viable industrial assets that can help future-proof operations.
Embodied AI: The integration of large language models and advanced spatial reasoning is granting robots the ability to “think” and act autonomously. This moves the field beyond mere mechanical motion into the realm of true adaptability, where robots can handle the “messy complexity” of our daily world (0:50–0:57).

As AI continues to bridge the gap in the “robotics puzzle” (2:47), the focus is shifting toward safe, intelligent, and general-purpose systems capable of performing diverse tasks rather than being limited to a single function.

Google DeepMind’s Gemini Robotics 2 marks a massive leap in physical AI by transitioning from isolated tabletop movements to unified whole-body intelligence for humanoid and bi-arm robots. Released as a next-generation three-model ecosystem, the stack seamlessly coordinates perception, logical reasoning, and physical motor control. [1, 2]

🧠 The Three-Model Architecture

Instead of relying on a single, rigid network, Gemini Robotics 2 splits its intelligence across three specialized layers: [3]

Gemini Robotics 2 (VLA Model): The primary Vision-Language-Action engine. It directly translates raw visual and text data into motor commands, driving the entire physical robot—from feet to fingertips—simultaneously. [1]
Gemini Robotics ER 2 (Embodied Reasoning): The “high-level brain”. This Vision Language Model (VLM) parses continuous video feeds, maps multi-step plans lasting several minutes, and coordinates complex problem-solving. [1, 4, 5]
Gemini Robotics On-Device 2: An optimized, low-latency action model designed to run locally. It can adapt to entirely unfamiliar robotic bodies in just a few hours using fewer than 200 physical data examples. [1, 3]

🦾 Breakthrough Core Capabilities

1. Unified Whole-Body Control [6]

Traditional AI models controlled a robot’s upper arms separately from its legs. Gemini Robotics 2 acts as a single cohesive nervous system. By dynamically calculating a humanoid’s center of gravity, it allows the robot to fluidly walk, reach, squat, and bend in cramped, real-world environments. In flagship tests with the Apptronik Apollo 2 Humanoid, the robot can process a command like “put the watering can on the bottom shelf,” walk over, pick it up, carry it across the room, and duck down to place it perfectly. [1, 2, 7, 8]

2. Advanced Finesse and Dexterity

The system masters complex endpoints, operating standard two-fingered grippers up to highly complex 22-degree-of-freedom hands (like the SharpaWave). It successfully demonstrates unprecedented physical finesse—such as tying intricate knots, unscrewing light bulbs, and sealing ziplock bags without explicit pre-programming. [1, 9]

3. Multi-Robot Collaboration

Robots running the Gemini stack can organically divide labor and hand off tasks. If a humanoid robot cannot reach a tight, low shelf, it can communicate via a shared semantic understanding to delegate the rest of the workflow to a nearby bi-arm or mobile robot. [5, 7, 10, 11]

📊 Performance and Real-World Metrics

According to Google DeepMind’s Internal Research Evaluations, the model breaks ground but highlights areas still requiring development: [2, 3]

Capability / Benchmark	Performance Metric	Notes / Current Bottlenecks
Light Bulb Unscrewing	92% Success Rate	Demonstrates elite full-body stabilization.
Moment Finding	91.3% Accuracy	Pinpoints exact video frames when events happen (e.g., bin is full).
Gripper Tasks (Franka Duo)	74.2% – 89.6% Success	Highly reliable on tight industrial packing.
Progress Tracking	57.4% Accuracy	Reliably catches mid-execution errors like spills or slips.
Complex Fiddly Tasks	32% – 44% Success	Actions like tying trash bags remain highly inconsistent.
Movement Speed	Needs Improvement	Physical executions are deliberate and slow.

🛠️ Developer Availability

The high-level reasoning model, Gemini Robotics ER 2, is widely available in preview via Google AI Studio and the Gemini API for building physical AI agents. The direct motor-control VLA and On-Device variants are currently restricted to early-access hardware partners. [1, 2, 4]