Gemini Robotics On-Device marks a major shift in how intelligent robots understand and interact with the physical world. Built as the most capable vision-language-action model to run entirely on hardware, it brings broad dexterity, rapid task adaptation, and robust performance to bi-arm robotic systems without relying on cloud infrastructure. This on-device variant extends the capabilities introduced earlier in Gemini Robotics 2.0, delivering real-world multimodal reasoning and practical manipulation skills directly on the robot. By operating independently of a data network, it addresses latency-sensitive applications and ensures resilience in environments where connectivity is unreliable or unavailable. To accelerate adoption and practical evaluation, a dedicated Gemini Robotics SDK is being released, enabling developers to test the system on their tasks and settings, explore the MuJoCo physics simulator, and adapt the model to new domains with as few as 50 to 100 demonstrations through a trusted tester program. This article provides a comprehensive, in-depth look at Gemini Robotics On-Device, its core capabilities, development tooling, and the implications for robotics across industries.
Gemini Robotics On-Device: A Self-Contained Vision-Language-Action Foundation for Bi-Arm Robots
Gemini Robotics On-Device represents a foundational leap in embedded robotics intelligence. It is engineered to function with minimal computational resources while delivering broad task generalization and dexterity capabilities that previously required larger, cloud-connected systems. The model embodies a design philosophy centered on rapid experimentation with dexterous manipulation, enabling researchers and engineers to iterate rapidly on new tasks within the robot’s own compute environment. By prioritizing on-device inference, the system minimizes round-trip latency and eliminates dependence on external data networks for most core operations, which is essential for real-time control and safe physical interaction.
From a modular perspective, Gemini Robotics On-Device inherits the strengths of its predecessor in the Gemini Robotics family, notably the emphasis on task generalization and dexterous manipulation. These capabilities are now optimized for local execution, preserving the model’s ability to interpret complex visual scenes, understand nuanced natural language instructions, and translate those instructions into precise, coordinated motor actions with two robotic arms. The on-device model maintains robust multimodal reasoning, allowing it to fuse perceptual cues with language guidance in a way that supports flexible and reliable manipulation across a variety of objects and contexts. In practical terms, this means a robot can interpret a spoken command or a written instruction, reason about possible approaches, and execute a sequence of manipulations—such as grasping, regrasping, coordinating both arms, and executing delicate operations—without needing external computation.
A core advantage of the On-Device approach is its resilience in real-world environments characterized by variability and noise. The model is designed to tolerate perceptual ambiguity, compensate for sensor imperfections, and maintain stable performance even when lighting, clutter, or occlusions challenge perception modules. The architecture is optimized to handle real-time perception-action loops, where perception updates continuously inform decision-making and motor control. By enabling the robot to operate offline, the system supports scenarios where network access is intermittent or non-existent, ensuring consistent behavior when cloud connectivity is unreliable or unavailable.
The on-device design also emphasizes safety and reliability. Because the robot can function independently, it can maintain a predictable operational profile even in degraded network conditions. This property reduces the risk of reliance on remote servers in critical applications, such as manufacturing lines, autonomous service robots, or field maintenance tasks where latency and connectivity can directly impact task success and safety margins. The model’s dexterity and instruction-following capabilities are paired with a lightweight, efficient inference engine that aligns with the computational budgets typical of embedded robotic platforms, enabling practical deployment on a wide range of hardware configurations.
In parallel with the core model, the Gemini Robotics SDK is positioned as a practical gateway for developers. The SDK provides streamlined access to evaluation tools, task environments, and integration hooks that simplify the process of testing Gemini Robotics On-Device on real-world tasks. Developers can explore model behavior, benchmark performance, and identify optimization opportunities tailored to their hardware, sensors, and control policies. The SDK also supports experimentation within MuJoCo, a physics simulator renowned for its realism in articulated-body dynamics, contact-rich manipulation tasks, and complex interactions with soft and rigid objects. This combination of on-device intelligence, robust real-world reasoning, and accessible tooling creates a comprehensive platform for advancing robotic autonomy in ways that are both powerful and implementable.
In terms of practical capabilities, Gemini Robotics On-Device brings strong general-purpose dexterity and task generalization to the robot’s on-board operation. It supports rapid experimentation with a range of dexterous manipulation tasks, enabling users to explore novel object-handling strategies without extensive computational overhead or cloud-based processing. The model is designed to be adaptable to new tasks through targeted fine-tuning, which can improve performance for specific domains or object sets while preserving generalization capacity. Importantly, the system is optimized for local execution with low-latency inference, ensuring that perceptual updates, decision-making, and motor commands occur within tightly bounded timeframes suitable for responsive robotic control.
The on-device model achieves broad visual, semantic, and behavioral generalization across diverse testing scenarios. It can follow natural language instructions with interpretability and stability, supporting a spectrum of commands from simple to complex. The model demonstrates high competence in highly dexterous tasks like unzipping bags or folding clothes, among others, all while operating directly on the robot. These demonstrations underscore the model’s ability to translate linguistic intent into precise, coordinated actions that require fine-grained control of both robotic arms and end-effectors. The combination of multimodal understanding, action planning, and motor execution on-device marks a notable milestone in reducing reliance on external compute for advanced manipulation.
To summarize this section, Gemini Robotics On-Device embodies a self-contained, bi-arm robotics foundation model that is engineered to run locally with minimal resource demands, while delivering advanced dexterity, generalization, and language-grounded control. The on-device paradigm reduces latency, enhances robustness, and simplifies deployment in environments where network access is uncertain. It also aligns with broader industry movements toward edge AI in robotics, where efficient, reliable, and adaptable perception-action systems are essential to scaling autonomous manipulation across sectors.
Core Capabilities: Dexterity, Generalization, and Language Alignment
Gemini Robotics On-Device is built to deliver a triad of core capabilities that together enable practical, reliable on-robot intelligence. First, dexterity refers to the model’s ability to plan and execute nuanced hand–eye coordination, multi-step manipulation sequences, and cooperative actions between two arms. This encompasses tasks that demand precision, timing, and adaptable grips—ranging from delicate grasping to multi-stage manipulation, where the robot must coordinate contact, force, and trajectory to achieve the desired outcome. The model’s dexterous control is designed to sustain performance across a variety of objects with different geometries, textures, and physical properties, including pliable materials that require compliant handling and dynamic adjustment of grips.
Second, task generalization is central to the model’s utility. The system is engineered to extend learned capabilities to new domains and object sets without requiring extensive reprogramming. It leverages structured representations of tasks, perceptual cues, and action affordances to generalize from prior experiences to unseen contexts. This generalization manifests in robust transfer to novel manipulation challenges, enabling the robot to abstract techniques from one setting and apply them to another with minimal additional data. The design anticipates real-world variability, where unforeseen objects, environments, and task sequences demand flexible adaptation.
Third, language-grounded behavior allows the robot to interpret instructions and align its actions with human intent. The model integrates visual input, semantic understanding, and motor planning to translate spoken or written commands into concrete manipulation policies. The language component supports task decomposition, sequencing, and error recovery, helping users specify objectives in an intuitive, natural form. This alignment with human communication is critical for collaborative robotics, where humans and machines share control responsibilities and require transparent, predictable interactions.
The On-Device architecture emphasizes efficient, low-latency inference, ensuring that perception, reasoning, and action generation occur within tight cycles. This is essential for tasks that require rapid decision-making, such as adapting to moving objects, compensating for sensor noise, or re-planning on the fly in response to environmental changes. Efficiency is achieved through a combination of model design choices, optimization techniques, and hardware-aware deployment strategies that keep computational demands within the limits of on-board resources while maintaining accuracy and reliability.
In practical deployment, these capabilities translate into a robot that can observe a scene, interpret a goal expressed in natural language, reason about possible manipulation strategies, and execute a sequence of precise motor commands. The system maintains a stable trajectory across steps, manages contact dynamics between the robot, objects, and environment, and handles contingencies through robust planning and recovery behaviors. The result is a cohesive, end-to-end capability set that supports a broad spectrum of real-world tasks without frequent reliance on external computation or cloud-based inference.
To illustrate the depth of these capabilities, consider scenarios such as unzipping a bag, where the robot must identify zipper teeth, grasp the zipper pull, coordinate both arms to maintain tension and alignment, and complete the action without damaging the bag or fabric. Similarly, folding clothes requires a series of coordinated motions, precise grasping at different points, and adaptive adjustments based on fabric drape and texture. Such tasks demonstrate the interplay of perception, planning, control, and language understanding, all operating on-device to deliver dependable performance.
The on-device model also supports rapid experimentation with dexterous manipulation. Engineers can introduce new task variants, objects, or constraints and observe how the model adapts in real time. The combination of dexterity, generalization, and language alignment is designed to lower the barrier to entry for researchers and practitioners, enabling a broader community to contribute to advancing robotic autonomy. This practical emphasis on real-world manipulation, rather than laboratory-only capabilities, reflects a commitment to translating sophisticated AI concepts into usable robotic tools that can impact manufacturing, service robotics, and consumer automation.
In summary, Gemini Robotics On-Device is a robust robotics foundation model for bi-arm systems, optimized for local operation and capable of broad dexterous manipulation, generalization across tasks, and natural-language-guided action. Its on-device inference reduces latency, enhances reliability in variable connectivity environments, and supports practical deployment across diverse settings. The model’s capabilities are reinforced by developer tooling and simulation environments that facilitate experimentation, domain adaptation, and rapid iteration, further strengthening the path from research to real-world impact.
On-Device Efficiency: Local Inference, Low Latency, and Resource-Aware Design
A defining feature of Gemini Robotics On-Device is its emphasis on efficiency, ensuring that high-performance perception, reasoning, and manipulation can occur entirely on the robot’s hardware. This implies a carefully engineered balance between model capacity, computational demands, memory footprint, and energy consumption. The design goal is to maximize task performance without pushing the device beyond its practical power envelopes, enabling continuous operation in field settings, industrial floors, and home environments where stability and endurance are essential.
Low-latency inference is central to on-device robotics. The model is structured to minimize the time between a perceptual input and motor output, a critical factor for successful manipulation in dynamic contexts. This entails optimizing the incremental update loop—perception updates, decision making, and motor commands—so that each cycle completes within a tightly bounded duration. The ability to react quickly to changes in the scene, moving objects, or human collaborators is a prerequisite for safe and effective human-robot interaction, and it is a core advantage of the on-device approach.
From a resource perspective, the model is designed to operate within the constraints of typical embedded hardware used in robotic platforms. This includes memory-efficient representations, streamlined attention mechanisms, and hardware-aware execution paths that exploit parallelism across cores and accelerators where available. The goal is to preserve performance across a spectrum of devices, from compact, energy-constrained platforms to more capable robotics cores, while maintaining consistent behavior and predictable power usage. The outcome is a system that can be deployed on a broad range of robots with varying sensor suites and actuator configurations, expanding accessibility for developers and end users alike.
Operational reliability on-device also benefits from robust fault tolerance and graceful degradation. In environments with sensor noise, occlusions, or mechanical variations, the model maintains functional integrity, prioritizing safe action and clear failure modes when necessary. This resilience is complemented by the ability to operate offline, ensuring that critical tasks can be completed even in the absence of network connectivity. When connectivity is available, the on-device system can operate in tandem with cloud-based services for tasks that require deeper analytics or long-term learning, but the core manipulation capabilities remain fully functional without external support.
The On-Device platform emphasizes modularity and upgradability. The architecture is designed for incremental improvements, enabling developers to roll out updates to the model, the software stack, or hardware acceleration without disrupting ongoing operations. This modularity also supports customization and fine-tuning for specialized use cases, allowing organizations to tailor dexterity, perception, and planning components to their specific objects, tools, or workflows. The combination of efficiency, low latency, and robust performance makes Gemini Robotics On-Device a practical solution for real-world deployment, where reliability and responsiveness are as important as raw capability.
In terms of real-world impact, efficiency enables longer operation times between charges, reduced maintenance overhead, and more predictable behavior in critical applications. On-Device operation reduces bandwidth consumption and lowers operational costs, particularly in scenarios with numerous robots deployed across a facility or in field deployments with limited connectivity options. The design philosophy emphasizes a sustainable balance between performance and practicality, ensuring the technology can be adopted at scale without compromising safety or reliability.
The integration with the MuJoCo physics simulator adds an important dimension to efficiency and testing. Developers can use MuJoCo to create, simulate, and refine manipulation tasks before deploying to physical robots. The simulator provides realistic contact dynamics, friction, and object interactions that mimic real-world behavior, enabling thorough validation of dexterous manipulation strategies in a risk-free environment. By evaluating model behavior in MuJoCo, engineers can iterate on task sequences, perception cues, and control policies to improve on-device performance, reduce unexpected failures, and accelerate the transition from simulation to deployment on the robot itself.
The SDK further enhances efficiency by providing streamlined workflows for evaluation and experimentation. Developers can sign up for the trusted tester program to access tools, documentation, and example environments that demonstrate how to run Gemini Robotics On-Device on their hardware. The SDK supports tasks with varied difficulty levels, enabling developers to quickly assess generalization capabilities, fine-tuning potential, and latency characteristics across different configurations. This ecosystem approach helps bridge research, development, and real-world deployment, aligning optimization goals with practical constraints and application requirements.
In conclusion, the efficiency and local inference capabilities of Gemini Robotics On-Device are central to its value proposition. By delivering low-latency, resource-aware performance on embedded hardware, the model enables reliable dexterous manipulation in a wide range of environments, including those with intermittent or zero connectivity. The combination of on-device reasoning, offline operation, and hardware-conscious design positions this system as a practical and scalable solution for real-world robotics applications, supported by simulation and developer tooling that encourages rapid experimentation and domain adaptation.
Developer Tools and Simulation: SDK Access, MuJoCo Testing, and Demonstration-Based Adaptation
A major component of Gemini Robotics On-Device is the developer ecosystem built around the Gemini Robotics SDK. The SDK is designed to lower barriers to entry for evaluating the on-device model, experimenting with tasks and environments, and adapting the system to specific domains with a minimal amount of demonstration data. Developers can access the SDK by joining the trusted tester program, which serves as a controlled pathway for early adoption and iterative refinement in collaboration with the model developers. Through the SDK, engineers gain access to a range of evaluation tools, sample environments, and practical guidelines for tuning performance on their hardware. The program is structured to ensure feedback loops are efficient and that improvements are informed by real-world use cases across diverse robotics platforms.
One of the key capabilities of the SDK is enabling testing in MuJoCo, a widely used physics simulation environment that provides high-fidelity models of articulated robots, joints, contacts, and interactions with a variety of objects. By simulating dexterous tasks in MuJoCo, developers can safely explore a broad spectrum of manipulation scenarios, quantify performance metrics, and identify corner cases that might not be easily reproduced in hardware-only testing. The MuJoCo environment allows researchers to model complex interactions, such as grasping, object reorientation, tool use, and constraining movements in cluttered spaces, all of which are valuable for refining on-device decision-making and control policies before real-world deployment.
The demonstration-based adaptation aspect of the SDK is particularly noteworthy. To tailor Gemini Robotics On-Device to a new domain, developers can provide 50 to 100 demonstrations that illustrate the target tasks, object varieties, constraints, and preferred success criteria. These demonstrations act as a compact but diverse data set that supports fine-tuning for improved performance in the specific domain while preserving the model’s generalization capabilities. The emphasis on a relatively small number of demonstrations is intentional, reflecting the goal of enabling rapid domain adaptation without the need for expansive data collection efforts. This approach strengthens the practical value of the SDK by making adaptation feasible in many use cases where data collection resources are limited.
The SDK’s tooling alternatives support a range of workflows that align with typical robotics development cycles. For example, engineers can start with a baseline evaluation in MuJoCo to understand how the on-device model responds to a given task sequence, object geometry, and environmental constraints. This initial assessment helps establish performance baselines, identify sensory or perception gaps, and determine whether additional fine-tuning or demonstrations are needed. As the development process progresses, the same workflows can be extended to hardware-in-the-loop testing, where simulated scenarios are gradually replaced with real-world trials, allowing for a controlled and staged validation process.
Documentation within the SDK emphasizes clarity, safety, and reliability. Developers will find guidelines on setting up the task scene, selecting appropriate objects, and configuring sensory inputs for faithful evaluation. The documentation also covers best practices for fine-tuning, including how to structure demonstrations, how to balance variation across demonstrations to improve generalization, and how to monitor for potential overfitting or regressive behavior. These resources help ensure that adaptation to new domains remains systematic, reproducible, and robust, preserving the integrity of the on-device inference while enabling practical improvements.
The developer ecosystem also prioritizes interoperability and extensibility. The architecture is designed to accommodate updates to the underlying model, perception modules, and control policies without destabilizing existing deployments. This openness supports ongoing innovation, allowing researchers to incorporate new perception backends, sensor modalities, or manipulation primitives as they emerge. The SDK’s design intent is to cultivate a vibrant community of developers who can share knowledge, reuse demonstration data where appropriate, and contribute to the collective advancement of on-device robotics.
In real-world terms, the combination of SDK accessibility, MuJoCo simulation, and demonstration-based adaptation creates a powerful loop for advancing on-device robotics. Engineers can prototype ideas quickly in a safe simulated environment, validate them against realistic physics, and then transfer proven improvements to the robot with a well-structured data and testing pipeline. The end result is a more efficient development cycle, shorter time-to-value for new domain deployments, and a stronger bridge between research concepts and tangible robotic capabilities.
Task Generalization, Fine-Tuning, and Domain Adaptation: Extending On-Device Capabilities to New Environments
A central promise of Gemini Robotics On-Device is its ability to generalize from learned experiences to unseen tasks and environments, with targeted fine-tuning enabling rapid domain adaptation. This section delves into how the model handles new domains, how fine-tuning is applied to improve performance, and how the system maintains generalization while specializing for particular use cases. The overarching goal is to deliver a robust, adaptable on-device intelligence that can be tailored to the realities of different robotic platforms, tasks, and operational contexts.
Generalization is achieved through training paradigms that emphasize diversified experiences and structured representations. The model is exposed to a wide range of manipulation scenarios during development, with varied object shapes, sizes, textures, and physical properties. It learns to recognize common manipulation primitives, such as grasping, releasing, reorienting, and aligning objects, and to compose these primitives into coherent action sequences that can adapt to new objects and configurations. The language-grounded control component helps the system interpret instructions that reference unfamiliar objects or tasks, enabling it to infer appropriate action plans by leveraging prior knowledge about similar tasks and interactions.
Fine-tuning for new domains uses a carefully designed process that balances specialization with the preservation of general capabilities. Developers can introduce a curated set of demonstrations that reflect the target domain’s objects, task sequences, and success criteria. These demonstrations should cover a spectrum of scenarios, including edge cases and variations in object appearance, lighting, and clutter. The fine-tuning process adjusts model parameters to improve performance on these domain-specific tasks while maintaining the model’s generalization strengths. The goal is to achieve meaningful improvements in task success rates, precision, and efficiency without compromising the model’s ability to generalize to unseen tasks.
Domain adaptation with a limited number of demonstrations—such as 50 to 100—offers practical advantages. It reduces data collection burdens and accelerates deployment timelines, especially in industries where new tasks arise frequently or environments change rapidly. The demonstration-based approach aligns with real-world workflows, where practitioners can curate compact yet representative example sets to guide the model’s adaptation. The result is a flexible programmatic path from broad capabilities to specialized competencies, enabling robots to excel in diverse settings—from manufacturing lines to service robotics, rehabilitation, and research laboratories.
From a functional perspective, on-device adaptation must contend with constraints such as memory, compute budgets, and real-time requirements. Fine-tuning strategies are designed to be lightweight and incremental, enabling updates to be applied without disrupting ongoing operations or requiring long downtime. This approach supports continuous improvement and rapid iteration, which are crucial for keeping up with evolving tasks, objects, and user expectations in dynamic environments. It also reduces the risk of catastrophic forgetting, where new domain-specific knowledge erodes performance on previously learned tasks.
The combination of task generalization, disciplined fine-tuning, and domain adaptation makes Gemini Robotics On-Device a versatile platform for real-world deployment. In practice, a robot could be deployed in a warehouse to handle a set of standard items while also accommodating a new product line with 50-100 demonstrations. The same robot could later be re-tuned for a home assistance scenario or a field maintenance task, using a compact demonstration set to achieve targeted improvements. This adaptability is essential for scaling robotics across multiple industries, where the same core technology can be repurposed to address a broad spectrum of tasks and environments.
To ensure safe and reliable adaptation, it is important to maintain rigorous evaluation during fine-tuning. The SDK and MuJoCo simulations provide controlled environments in which performance metrics, safety constraints, and failure modes can be studied before deploying to physical hardware. This disciplined approach helps prevent unintended consequences, such as overfitting to narrow demonstrations or introducing unstable behaviors when confronted with novel objects or contexts. By combining offline evaluation, controlled simulations, and incremental on-device updates, developers can confidently expand the robot’s capabilities while preserving reliability and safety.
In summary, the task generalization, fine-tuning, and domain adaptation capabilities of Gemini Robotics On-Device enable practical, scalable customization for a wide range of environments. The ability to enhance on-device performance with a small, carefully chosen set of demonstrations supports rapid deployment and ongoing improvement, while preserving generalization to new tasks. This balance between specialization and generalization is a key enabler of real-world applicability, ensuring that on-device autonomy remains robust as tasks evolve and new domains emerge.
Applications, Implications, and Real-World DeploymentScenarios
Gemini Robotics On-Device is designed to empower a broad spectrum of applications by delivering robust, on-board vision-language-action capabilities that can operate without cloud access. In industrial settings, bi-arm robots with strong dexterity and language-aligned control can perform complex assembly, quality inspection, material handling, and repetitive tasks with a higher degree of autonomy and precision. The on-device nature reduces latency, enabling faster cycle times, more responsive human-robot collaboration, and greater resilience during network outages. These attributes translate into tangible efficiency gains, improved safety margins, and the potential to reconfigure production lines quickly in response to changing product mixes or demand patterns.
In service robotics, the ability to operate offline and respond to natural language commands in real time opens up opportunities for patient care, hospitality, and assisted living contexts. A robot can understand a user’s instructions, interpret environmental cues, and carry out multi-step manipulation tasks without relying on remote servers. Such capabilities support more natural human-robot interactions, where clarity of intent and timely responses contribute to a more seamless and intuitive user experience. The capacity to fold clothes, unzip bags, or arrange items—tasks that can be representative of everyday household activities—demonstrates the practical relevance of on-device dexterity for consumer robotics.
Field maintenance and remote operations represent another important domain. On-device inference ensures that robots can operate in remote or hazardous environments where connectivity is limited or impractical. For example, inspection drones or robotic assistants in mining, energy, or infrastructure sectors can benefit from immediate perception-action loops, enabling rapid decision-making and manipulation tasks without reliance on external networks. The ability to perform tasks with minimal data transfer reduces bandwidth requirements and enhances security and privacy by limiting data exposure in transit.
Security, privacy, and reliability considerations accompany these deployments. On-device processing minimizes data transfer to cloud services, reducing potential exposure of sensitive information. The on-board model can be designed with privacy-preserving aspects, such as local data handling, secure bootstrapping, and controlled update mechanisms. Reliability is enhanced by offline operation, but it also demands robust fail-safe behavior and transparent interpretability. Designers and operators must ensure that the robot can gracefully degrade or halt operations if the perception or control system encounters uncertain conditions, incorporating explicit safety protocols and human-in-the-loop options where appropriate.
From an economic perspective, the ability to adapt to new domains with a relatively small number of demonstrations can lower the total cost of ownership for robotic systems. Organizations can deploy base capabilities across multiple tasks and then refine per-domain performance with targeted demonstrations, reducing the need for large, expensive datasets or extensive reprogramming. This approach aligns well with iterative development strategies and allows for rapid prototyping, testing, and rollouts across departments and sites. The SDK and MuJoCo tooling support this process by providing a repeatable framework for validation, benchmarking, and performance-driven optimization.
Looking ahead, Gemini Robotics On-Device foreshadows a future in which on-device intelligence becomes a mainstream component of robotic systems. By combining dexterous manipulation, robust generalization, and natural language-guided control with efficient on-board execution, this technology has the potential to accelerate robotics adoption across industries, enabling more agile manufacturing, more capable service robots, and more autonomous tools for field operations. The integration of developer tooling, simulation environments, and demonstration-driven adaptation supports a practical pipeline for continued innovation, bridging the gap between research breakthroughs and real-world impact.
The broader implications include a shift in how robotic autonomy is conceived and delivered. On-device models encourage decentralized computation, closer integration with perception hardware, and more intimate coupling between AI reasoning and physical control. This evolution can foster new workflows, collaboration patterns, and business models that emphasize edge AI, hardware-software co-design, and end-to-end optimization from perception to action. As organizations explore these capabilities, they will also need to consider governance, safety standards, and regulatory compliance to ensure that advanced robotic systems operate responsibly and predictably in diverse contexts.
Conclusion
Gemini Robotics On-Device represents a comprehensive, on-board vision-language-action solution designed for bi-arm robots, delivering strong dexterity, broad task generalization, and natural-language-aligned control while operating locally with low latency. Built upon the progress demonstrated by Gemini Robotics and Gemini 2.0, this on-device variant prioritizes latency-sensitive performance and resilience in environments with intermittent or zero network connectivity. The accompanying Gemini Robotics SDK, together with MuJoCo simulation capabilities and demonstration-based adaptation, provides a practical, scalable pathway for developers to evaluate, refine, and deploy on-device robotic capabilities across a range of domains. By combining efficient on-device inference, robust generalization, and accessible tooling, Gemini Robotics On-Device aims to accelerate the practical adoption of autonomous manipulation in industrial, service, and field environments, while preserving safety, privacy, and reliability as core design principles.
