The definitive architecture for general-purpose robotics is emerging, separating high-level, cloud-based task planning from low-latency, local motor control. Google DeepMind has advanced this paradigm with Gemini Robotics 2, a system designed to move robotic capabilities beyond repetitive, scripted actions. A single Gemini Robotics 2 model checkpoint directly controls three distinct physical hardware configurations, including full-body Apptronik Apollo 2 humanoids and Franka Duo grippers, achieving up to 92% success on complex multi-finger tasks.
This release introduces a tiered approach, featuring a vision-language-action (VLA) model for motor control, Gemini Robotics ER 2 for high-level embodied reasoning, and Gemini Robotics On-Device 2 for local execution. This modularity aims to enable robots to adapt to varied environments and tasks with greater flexibility, signaling a shift towards more versatile autonomous systems.
Cross-Embodiment Generalization and Rapid Adaptation
According to Google DeepMind, a single Gemini Robotics 2 model checkpoint demonstrated direct control over three distinct physical embodiments, including Apptronik’s Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and a Franka Duo parallel gripper platform. This cross-embodiment capability allows the same underlying intelligence to drive diverse hardware, from humanoid forms to industrial manipulators.
The system demonstrates impressive performance on specific tasks. Multi-finger dexterity on Apollo 2 with SharpaWave hands achieved a 92% success rate in unscrewing a lightbulb. For the Franka Duo with a Robotiq gripper, precise insertion tasks saw an 89.6% success rate, while diverse tool kitting reached 78.9%, and general pick and place tasks achieved 74.2%.
Gemini Robotics On-Device 2, which inherits motion transfer techniques from Gemini Robotics 1.5, accepts text, images, and robot proprioception as numerical values for inputs, and produces robot actions as numerical values for outputs. This model demonstrated rapid adaptation to new robot bodies post-training using fewer than 200 examples. New bi-arm embodiments can be adapted by On-Device 2 within a few hours, a process that remains consistent despite variations in shape, sensors, or degrees of freedom. This significantly improved success rates on the SO101 platform from 6.7% (with On-Device 1) to 53.3%, and on the Dexmate platform from 24.4% at the start to 75.6% post-training, compared to On-Device 1 which improved from 13.3% to 33.3%.
Navigating Dexterity Limits and Fragmented Access
Despite advancements, the operational drop-off in fine dexterity highlights persistent real-world motor execution limits. Multi-finger dexterity performance on Apollo 2 with SharpaWave hands ranged from the high of 92% for unscrewing a lightbulb to significantly lower scores, such as 44% for tying a trash bag, 40% for ziplock operation, 36% for screwing a lightbulb, and 32% on dustpan operation. Moreover, Gemini Robotics On-Device 2 faces documented challenges when generalizing to out-of-distribution tasks and in its ability to control high-degree-of-freedom robots.
The architectural separation also comes with a tiered access strategy. Gemini Robotics ER 2, built on Gemini 3.5 Flash with a 128k context window, is available via public preview on the Gemini API and Google AI Studio, with its launch link using the model string gemini-robotics-er-2-preview. It is also in private preview on the Gemini Enterprise Agent Platform. This high-level embodied reasoning model emits text output up to 64K tokens and integrates with Gemini Live API via a bidirectional streaming endpoint. It achieves 57.4% accuracy in video progress classification, which evaluates video feeds divided into 5 progress levels, and 91.3% accuracy in critical event moment finding, with a 0.96 second mean absolute distance. The system’s general instrument reading capabilities were evaluated across 10 instrument types, encompassing devices such as digital displays, linear scales, rulers, and liquid thermometers.
In contrast, the Gemini Robotics 2 (VLA) model is available to early-access partners, while On-Device 2 is available to Trusted Testers. This fragmented access means developers relying on the publicly available ER 2 for high-level planning remain dependent on network latency and multi-tiered tool execution abstractions, rather than having immediate access to end-to-end, locally autonomous control stacks. Google DeepMind also released ASIMOV-Agentic, a dedicated robotics safety benchmark, on Hugging Face under the CC-BY-4.0 license, indicating a focus on responsible deployment.
📊 Key Numbers
- Apollo 2 + SharpaWave hands (Unscrew lightbulb) success rate: 92%
- Apollo 2 + SharpaWave hands (Dustpan operation) success rate: 32%
- Apollo 2 + SharpaWave hands (Screw lightbulb) success rate: 36%
- Apollo 2 + SharpaWave hands (Tie trash bag) success rate: 44%
- Apollo 2 + SharpaWave hands (Ziplock operation) success rate: 40%
- Franka Duo (Precise insertion tasks) success rate: 89.6%
- Franka Duo (Diverse tool kitting) success rate: 78.9%
- Franka Duo (General pick and place) success rate: 74.2%
- Apollo 2 + Inspire hands (Pick up from shelf) success rate: 76.3%
- Apollo 2 + Inspire hands (Pick up from table) success rate: 68.4%
- Apollo 2 + Inspire hands (Pick up from floor) success rate: 45.7%
- Gemini Robotics ER 2 (Video progress classification accuracy): 57.4%
- Gemini Robotics ER 2 (Critical event moment finding accuracy): 91.3%
- SO101 (On-Device 2) post-training task success: 53.3% (vs 6.7% for On-Device 1)
- Dexmate (On-Device 2) post-training task success: 75.6% (vs 33.3% for On-Device 1)
- Gemini Robotics ER 2 context window: 128k tokens
- Gemini Robotics ER 2 text output: up to 64K tokens
- Moment finding accuracy mean absolute distance: 0.96 seconds
- Video progress classification levels: 5 (0-20% through 80-100%)
- General instrument reading types tested: 10
🔍 Context
This announcement addresses the persistent challenge of creating truly general-purpose robots that can operate reliably across diverse, unstructured environments without extensive re-programming. By decoupling high-level reasoning from low-level motor control, Google DeepMind aims to overcome the limitations of traditional, tightly coupled robot control systems that often require bespoke engineering for each new task or hardware configuration. This approach accelerates the trend towards more flexible and adaptable robotic systems, moving beyond the confines of factory floors and into more dynamic settings.
The release of Gemini Robotics 2, with its distinct VLA, ER 2, and On-Device 2 models, directly responds to the need for more robust and generalizable robotic intelligence. The system’s ability to control multiple robot embodiments from a single model checkpoint, as demonstrated with Apptronik’s Apollo 2 and the Franka Duo, represents a significant step towards this goal. The availability of Gemini Robotics ER 2 via public preview on the Gemini API and Google AI Studio allows developers to begin experimenting with the high-level reasoning capabilities, while the gated access for VLA and On-Device 2 models suggests a controlled rollout for the more critical, real-time control components.
💡 AIUniverse Analysis
Our reading: The genuine advance here lies in the architectural decoupling of high-level task planning from low-latency motor control, enabling a single model checkpoint to drive diverse physical embodiments. This cross-embodiment generalization, particularly the rapid adaptation of Gemini Robotics On-Device 2 to new robot bodies using fewer than 200 examples and within hours, is a critical mechanism for scaling robotics beyond single-purpose machines. It promises a future where the same AI brain can inhabit different robot bodies, significantly reducing development overhead for new hardware.
However, the shadow cast by this announcement is the stark operational drop-off in fine dexterity, with success rates plummeting to 36% for screwing a lightbulb and 32% for dustpan operation on the Apptronik Apollo 2 with SharpaWave hands. This highlights that even with advanced reasoning, continuous physical interaction remains severely constrained by real-world motor execution limits. Furthermore, by placing the real-time VLA and On-Device models behind gated early-access programs while only opening the ER 2 reasoning tier, Google DeepMind leaves developers dependent on network latency and multi-tiered tool execution abstractions, rather than providing end-to-end, locally autonomous control stacks. This fragmentation could hinder the rapid iteration and deployment needed for truly general-purpose, real-world robotics. For this to matter in 12 months, Google DeepMind would need to demonstrate significant improvements in fine motor control robustness and provide more integrated, low-latency access to the full control stack.
⚖️ AIUniverse Verdict
👀 Watch this space. While the cross-embodiment control is a significant architectural step, the documented limitations in fine dexterity and the fragmented access tiers for critical control models mean its real-world impact remains unproven.
🎯 What This Means For You
Founders & Startups: Robotics startups can build sophisticated high-level task planners using standard API tool-calling interfaces without needing to train custom reasoning vision-language models.
Developers: Developers can access the gemini-robotics-er-2-preview model via Google AI Studio and utilize sample integration code in the official robotics-samples repository for video-based task tracking.
Enterprise & Mid-Market: Industrial enterprises can deploy heterogeneous robot fleets—combining humanoids and bi-arm manipulators—that collaborate through shared semantic task transfers without replacing existing physical platforms.
General Users: Everyday consumers gain a realistic benchmark for home robotics, as whole-body humanoids demonstrate multi-room chore capabilities despite remaining limited by slow movement speeds.
⚡ TL;DR
- What happened: Google DeepMind released Gemini Robotics 2, enabling a single model checkpoint to control diverse robot hardware, including humanoids and grippers, with tiered access to its reasoning and control components.
- Why it matters: This system advances general-purpose robotics by decoupling high-level task planning from low-latency motor control, allowing for rapid adaptation to new robot bodies and multi-robot collaboration.
- What to do: Monitor the progress of fine motor control capabilities and the availability of the full, low-latency control stack beyond the publicly accessible reasoning model.
📖 Key Terms
- Vision-Language-Action (VLA)
- A model type that processes visual and linguistic inputs to generate robot actions for motor control.
- Gemini Robotics ER 2
- Google DeepMind’s model for high-level embodied reasoning, built on Gemini 3.5 Flash, available via public preview.
- ASIMOV-Agentic
- A dedicated robotics safety benchmark released by Google DeepMind on Hugging Face.
- SharpaWave hand
- A specific type of multi-finger robotic hand used with Apptronik’s Apollo 2 humanoid platform.
- Embodied Reasoning (ER)
- The capability of an AI system to understand and reason about its physical environment and its own body to plan and execute complex tasks.
Analysis based on reporting by MarkTechPost. Original article here.
🔗 Sources Consulted
- Primary source: MarkTechPost
- Additional source consulted: github.com
Figures and claims in this article were checked against the documents listed above. Items that could not be traced to them were removed before publication.