Introduction: WAIC 2026 and the Coming-of-Age of Humanoid Robots
The WAIC 2026 summit offered a panoramic showcase of humanoid robotics. More than 200 robotics enterprises and over 300 robot models gathered at the exhibition venue. Demonstrations spanned versatile shape-shifting robots from Sany Heavy Industry, and intelligent robots soldering 2-millimeter resistors onto circuit boards. Past robotics showcases prioritized spectacular mobility performances, including aerial hovering and mechanical movement displays. Yet manufacturers are shifting their core goal: proving robots are capable of practical work, instead of merely moving.
While robot hardware has advanced rapidly, enabling robots to walk, jump and stretch their limbs freely, hardware is no longer the primary bottleneck. The real challenge lies in the robot’s “brain”. Embodied machines need to interpret unstructured chaotic environments, make sequential action decisions, and coordinate dozens of joints consistently over long-running tasks. This cognitive gap has long restricted widespread commercial deployment of humanoid robots.
One week after the closing of WAIC 2026, Google DeepMind officially released Gemini Robotics 2, a general-purpose robot model built around embodied full-body intelligence. This system targets robots of diverse physical forms. Human operators can complete sequential tasks such as fetching kettles and sealing bags naturally, yet robots face far higher control complexity when chaining a series of connected movements. The more physical components a robot carries, the harder coordinated control becomes. This article dissects the architecture, benchmark results, core advantages and unresolved limitations of Gemini Robotics 2, alongside the broader industry outlook for embodied intelligence.
Core Composition of Gemini Robotics 2
Google’s latest robotics suite consists of three coordinated model modules: Gemini Robotics ER 2, Gemini Robotics 2 VLA model, and the On-Device variant.
- Gemini Robotics ER 2: Responsible for environmental comprehension, long-term task planning, and natural language interaction with human operators.
- Gemini Robotics 2 VLA (Vision-Language-Action): Translates visual observations and language instructions directly into continuous, high-frequency robotic motions.
- Gemini Robotics On-Device: The lightweight local deployment variant. It can adapt to new robot hardware platforms with fewer than 200 demonstration samples and only several hours of training data.
The core design philosophy is unified model coordination. The system synchronously controls humanoid robots’ legs, torsos, arms and dexterous fingers. It enables robots to reason about objectives while executing physical operations simultaneously, breaking the traditional sequential workflow of “planning first, executing later”.
Capability Demonstrations and Measured Performance Data
The prior generation of Gemini Robotics concentrated mainly on upper-body manipulation for desktop tasks. Gemini Robotics 2 expands control scope to full-body motion. Official demo footage features the Apptronik Apollo humanoid robot performing complete workflows: walking toward water dispensers, lifting kettles, and picking up gloved balls. All motions run on the robot’s native hardware without playback prerecording. The model treats the entire physical body as a coupled action space, and a single model checkpoint supports multiple robot hardware platforms.
Google openly published quantitative success rates from standardized testing scenarios:
- Object retrieval from desktops: 68.4%
- Object retrieval from flat ground: 45.7%
- Object retrieval from storage shelves: 76.3%
Ground-level manipulation registers the lowest success rate among the three scenarios. Google acknowledges that the movement speed of Apollo remains suboptimal and requires further iterative optimization.
Industry analysts name the challenge “the last few centimeters” problem. Full-body navigation enables robots to approach target locations and stabilize stances. However, fine-grained interaction ultimately determines whether robots can reliably enter household and factory environments. Demo footage shows Apollo utilizing five-finger SharpaWave grippers to screw light bulbs, reseal packaging bags and tie garbage bags. Franka Duo robotic arms assemble hardware components into toolboxes. These workflows demand sustained physical contact, bilateral coordination and adaptive deformation.
Compared with simple pick-and-place demos, Gemini Robotics 2 can generate continuous motion sequences directly from visual and linguistic inputs. Success rates stay relatively high for tool assembly and precision fitting. Meanwhile, multi-finger manipulation tasks including bulb installation and garbage bag tying maintain lower success metrics. The data reveals a clear boundary: large embodied models can handle complex gross hand movements, yet high-precision dexterous manipulation remains inferior to human capability.
Multi-Robot Collaboration Exploration
Google released footage demonstrating cooperative task execution between the Apollo humanoid robot and Franka Duo robotic arms. The two machines operate on separate model instances, distributing assembly workflows via high-level reasoning. The ER 2 module analyzes video feedback to track task progress and decide whether reattempts are required.
This workflow differs fundamentally from rigid pre-programmed mechanical scripts. Google aims to enable heterogeneous robot hardware to hand over tasks just like human teams. Nevertheless, the official demonstration takes place under tightly controlled lab conditions. Stable multi-robot coordination inside real production workshops still requires extensive verification before industrial adoption.
Abandoning the Traditional “Big Brain + Small Brain” Control Architecture
A widespread industrial paradigm for humanoid robots splits the control stack into two layers. Large foundation models handle high-level reasoning and task planning, while lightweight specialized controllers manage low-level joint motion. This layered framework carries practical advantages, yet it introduces inherent boundary conflicts. Inconsistencies frequently emerge between high-level plans and low-level physical execution, accumulating errors and triggering abrupt motion halts.
Google chooses an alternative route with ER 2: high-level unified control covering the entire robot body. This design reduces motion fractures caused by communication and parsing mismatches between hierarchical modules. It creates a clearer boundary: high-level logic runs end-to-end across the physical robot body. Low-level motion control and safety constraints are not packed into a single monolithic model. This architectural direction is not exclusive to Google; Figure AI and NVIDIA Isaac also pursue comparable technical exploration.
Ongoing Challenges Facing Embodied Intelligence
The research community building embodied intelligence models converges on shared targets: supporting diverse robot hardware, handling extended task sequences, and minimizing reliance on specialized labeled datasets. NVIDIA, Physical Intelligence and Lingdong Intelligence have all published relevant research outcomes.
The On-Device adaptation capability of Gemini Robotics 2 stands out as a valuable advancement. Nonetheless, robot control poses harder challenges than generic large language models. NVIDIA researchers have noted persistent obstacles in multi-task language following and distributed externalization. Low success rates for five-finger manipulation confirm stability deficits when one model controls diversified robot hardware. Real-world deployment will confront far more unforeseen interference than closed laboratory environments.
Developers operating distributed multi-model robot systems frequently manage complex API routing and access permission scheduling. Platforms such as 4sapi deliver streamlined gateway orchestration to simplify traffic management across heterogeneous model endpoints for robotics research teams.
Long-Term Prospects for Embodied Intelligence and the World Robotics Conference 2026
The most meaningful breakthrough brought by Gemini Robotics 2 lies in extending embodied model capabilities beyond desktop upper-body manipulation. The system incorporates full-body kinematics as core input variables. Mature chips, joint actuators and integrated robot hardware showcased at WAIC supply the physical foundation for AI models. Google’s technical results prove foundation models are starting to treat robot bodies as integral parts of the reasoning system.
Even with mature embodied intelligence, robots still face multiple barriers before widespread household adoption. From August 19 to August 23, the 2026 World Robotics Conference (WRC) will be held in Beijing. The conference theme “Human-Machine Coexistence, Shared Prosperity” will gather more than 300 exhibitors, representing a 36% year-on-year increase. Over 2,000 exhibits and more than 150 newly launched products will be unveiled. Approximately 10,000 contestants will participate in robotics challenge tournaments. Media teams will conduct full on-site coverage of the WRC exhibition.
The conference creates a vital observation window for tracking embodied intelligence progress. Industry practitioners will focus on a core question: can embodied intelligence models like Gemini Robotics 2 deliver stable, usable value within daily household and industrial scenarios?
Conclusion
Gemini Robotics 2 marks a critical milestone for Google DeepMind’s embodied robotics research. The three-module system realizes full-body coordinated control, supports cross-hardware migration with limited sample data, and delivers measurable benchmark improvements on common manipulation tasks. Test data simultaneously exposes clear limitations: low precision dexterous manipulation success rates, imperfect multi-robot coordination stability, and remaining gaps between lab demonstrations and chaotic real-world environments.
The paradigm of separated “high-level planning and low-level control” is gradually being re-examined. End-to-end unified embodied models represent a mainstream exploratory direction. Moving forward, researchers must balance model generalization, motion stability, safety constraints and deployment costs. Future iterations of Gemini Robotics and competing embodied models will determine the timeline for humanoid robots to step out of laboratories and enter ordinary work and living spaces.
