Gemini Robotics 1.5: DeepMind’s ER↔VLA Stack Brings Agentic Robots to the Real World

Gemini Robotics 1.5: DeepMind’s ER↔VLA Stack Brings Agentic Robots to the Real World

Estimated reading time: 6 minutes

• Modular Architecture: Gemini Robotics 1.5 introduces a novel ER↔VLA stack, separating high-level embodied reasoning (Gemini Robotics-ER 1.5) from low-level visuomotor control (Gemini Robotics 1.5 VLA).
• Motion Transfer: Skills learned on one robot can be transferred zero-shot to diverse, heterogeneous platforms, significantly reducing data collection and deployment costs.
• Agentic Capabilities: The system enables long-horizon, multi-step autonomy with tool-augmented planning, allowing robots to leverage external information (e.g., web search, APIs) for context-aware decision-making.
• Enhanced Reliability: Explicit “think-before-act” traces in the VLA improve planning, error recovery, and overall task completion rates for complex operations.
• Safety-First Design: DeepMind integrates layered safety controls, policy-aligned dialog, safety-aware grounding, and advanced evaluation techniques like auto red-teaming to ensure secure and reliable robot operation.

• Key Takeaways
• The ER↔VLA Architecture: A New Paradigm for Embodied AI
• Unlocking Versatility: Motion Transfer and Cross-Robot Learning
• Real-World Agentic Intelligence: Applications and Robust Safety Measures
• Actionable Steps for Robotics Innovators
• Conclusion
• Frequently Asked Questions

The dream of truly intelligent robots, capable of adapting to complex, unpredictable real-world environments, has long been a frontier in AI. Traditional robotics often grapples with the challenge of building systems that can not only execute precise motions but also understand their surroundings, plan dynamically, and learn new skills efficiently. DeepMind’s latest innovation, Gemini Robotics 1.5, marks a significant leap towards realizing this vision, introducing a modular architecture designed to tackle these intricate demands head-on.

“Can a single AI stack plan like a researcher, reason over scenes, and transfer motions across different robots—without retraining from scratch? Google DeepMind’s Gemini Robotics 1.5 says yes, by splitting embodied intelligence into two models: Gemini Robotics-ER 1.5 for high-level embodied reasoning (spatial understanding, planning, progress/success estimation, tool-use) and Gemini Robotics 1.5 for low-level visuomotor control. The system targets long-horizon, real-world tasks (e.g., multi-step packing, waste sorting with local rules) and introduces motion transfer to reuse data across heterogeneous platforms.”

This revolutionary approach addresses fundamental limitations of prior embodied AI systems, paving the way for more robust, versatile, and context-aware robotic agents. By clearly delineating the responsibilities of reasoning and control, DeepMind is building a foundation for robots that can truly “think” and “act” in harmony within dynamic environments.

The ER↔VLA Architecture: A New Paradigm for Embodied AI

At the heart of Gemini Robotics 1.5 lies its innovative two-model architecture, known as the ER↔VLA stack. This strategic split of embodied intelligence into distinct, yet interconnected, components is crucial for handling the complexity of real-world tasks. It tackles the shortcomings of earlier end-to-end Vision-Language-Action (VLA) models, which often struggled with robust planning, verifying task success, and generalizing across different robotic platforms.

Gemini Robotics-ER 1.5: The Reasoner

The first component is Gemini Robotics-ER 1.5, often referred to as the “reasoner” or “orchestrator.” This is a multimodal planner designed for high-level embodied reasoning. It processes a rich array of inputs, including images, video, and even optional audio, to build a comprehensive understanding of the scene. The ER model grounds references using 2D points, meticulously tracks task progress, and estimates the likelihood of success. Critically, it can invoke external tools like web search or local APIs to fetch vital constraints or information before issuing detailed sub-goals. Developers can access Gemini Robotics-ER 1.5 via the Gemini API in Google AI Studio, making its advanced planning capabilities more accessible.

Gemini Robotics 1.5: The VLA Controller

The second pillar is Gemini Robotics 1.5, functioning as the low-level Vision-Language-Action (VLA) controller. This model’s primary role is to translate instructions and sensory perceptions into precise motor commands. A key feature of this VLA is its ability to produce explicit “think-before-act” traces. These intermediate reasoning steps decompose long, complex tasks into manageable short-horizon skills, significantly enhancing the robot’s ability to plan, adapt, and recover from errors during execution. Currently, the VLA controller’s availability is limited to selected partners during its initial rollout phase.

This modularity—isolating deliberation (scene reasoning, sub-goaling, success detection) from execution (closed-loop visuomotor control)—brings substantial benefits. It improves interpretability by making internal thought processes visible, aids in faster and more reliable error recovery, and dramatically boosts the long-horizon reliability of robotic operations.

Unlocking Versatility: Motion Transfer and Cross-Robot Learning

One of Gemini Robotics 1.5’s most groundbreaking contributions is its capability for Motion Transfer (MT). This innovation directly addresses a long-standing challenge in robotics: the need for extensive, robot-specific data collection and retraining for every new platform or task. Motion Transfer allows skills learned on one robot to be transferred zero-shot to another, vastly accelerating deployment and reducing development costs.

DeepMind achieved this by training the VLA on a unified motion representation, meticulously built from data across a diverse range of heterogeneous robot platforms, including ALOHA, bi-arm Franka, and Apptronik Apollo. This means that a single VLA checkpoint can reuse skills across entirely different embodiments. Such cross-embodiment prior knowledge not only reduces the per-robot data burden but also helps narrow the notorious sim-to-real gap, making lab-developed skills more robust in physical environments.

The research team performed rigorous quantitative evaluations, showcasing compelling results from controlled A/B comparisons on real hardware and aligned MuJoCo scenes:

• Superior Generalization: Robotics 1.5 significantly outperforms previous Gemini Robotics baselines across instruction following, action generalization, visual generalization, and task generalization on all three platforms tested.
• Effective Zero-shot Cross-Robot Skills: Motion Transfer yields measurable gains in both progress and success when skills are transferred across different robot embodiments (e.g., Franka to ALOHA, or ALOHA to Apollo), demonstrating more than just partial progress.
• “Thinking” Enhances Acting: Activating the VLA’s thought traces demonstrably increases long-horizon task completion rates and stabilizes mid-rollout plan revisions, proving the value of explicit reasoning during execution.

These results confirm that the ability to transfer skills and the explicit reasoning steps are not just theoretical advantages but lead to tangible improvements in real-world robot performance.

Real-World Agentic Intelligence: Applications and Robust Safety Measures

Gemini Robotics 1.5 represents a significant paradigm shift from “single-instruction” robotics towards genuinely agentic, multi-step autonomy. This means robots are no longer confined to simple, predefined actions but can engage in complex sequences, use external tools, and learn across different platforms. This capability set has profound implications for both consumer and industrial robotics, enabling machines to handle more sophisticated and diverse tasks than ever before.

Tool-Augmented Planning and Context-Aware Autonomy

A prime example of this agentic intelligence is the tool-augmented planning enabled by Gemini Robotics-ER 1.5. The ER model can invoke external tools, such as performing a web search to fetch real-time data or accessing local APIs for specific regulations, to inform and condition its plans. Consider a robotic arm tasked with sorting household waste. Traditionally, this would require extensive training for each new object or specific rule. With Gemini Robotics 1.5, the ER model can leverage external APIs to fetch local recycling regulations—for instance, ‘plastic bags are not recyclable in this city’—and incorporate these into its high-level plan. The VLA then executes the precise movements, observing its own progress and adjusting if an item is initially misidentified, demonstrating robust, context-aware autonomy.

Robust Safety Measures

DeepMind places a strong emphasis on safety and robust evaluation. The research team highlights a system of layered controls designed to mitigate risks. These include policy-aligned dialog and planning, safety-aware grounding (ensuring the robot doesn’t point to hazardous objects), and adherence to low-level physical limits. Furthermore, DeepMind has expanded its evaluation suites, incorporating ASIMOV/ASIMOV-style scenario testing and auto red-teaming techniques. These advanced methods are crucial for eliciting edge-case failures, probing for hallucinated affordances (believing an object has a certain use when it doesn’t), or identifying nonexistent objects before any physical actuation occurs, ensuring the robot operates safely and reliably.

The competitive and industry context for Gemini Robotics 1.5 is vast. Its capabilities in multi-step autonomy, explicit web/tool use, and cross-platform learning are highly relevant to established robotics vendors and emerging humanoid platforms. Early partner access is strategically centered on these key players, indicating a clear path toward real-world integration.

Actionable Steps for Robotics Innovators

• Engage with the Gemini API: For those developing high-level robot intelligence, explore Gemini Robotics-ER 1.5 through the Gemini API in Google AI Studio. Experiment with its multimodal planning, scene reasoning, and tool-use capabilities to orchestrate complex tasks.
• Design for Modular Cognition: Consider adopting the ER↔VLA paradigm in your own robot agent designs. Separating high-level reasoning from low-level control can significantly improve interpretability, error recovery, and the robustness of your long-horizon robotic applications.
• Investigate Cross-Platform Data Reuse: For organizations managing diverse robot fleets, DeepMind’s Motion Transfer highlights a powerful avenue for efficiency. Explore how unifying motion representations could enable zero-shot skill transfer across your heterogeneous robot platforms, reducing data collection overhead and accelerating deployment.

Conclusion

Gemini Robotics 1.5 represents a transformative moment in the field of embodied AI. By operationalizing a clean separation of embodied reasoning (ER) and low-level visuomotor control (VLA), DeepMind has engineered a robust, interpretable, and highly adaptable system. The introduction of Motion Transfer capabilities drastically reduces the data burden typically associated with training new robot skills, enabling efficient reuse of learned behaviors across heterogeneous platforms.

Furthermore, the system’s capacity for “think-before-act” control and tool-augmented planning empowers robots to tackle long-horizon, complex tasks with unprecedented reliability and context-awareness. Coupled with DeepMind’s comprehensive approach to safety through layered controls and advanced evaluation suites, Gemini Robotics 1.5 is not just an advancement in technology; it’s a significant stride towards bringing truly agentic, versatile, and safe intelligent robots into our real world.

Frequently Asked Questions

• What is Gemini Robotics 1.5?

  Gemini Robotics 1.5 is DeepMind’s latest innovation in embodied AI, featuring a modular ER↔VLA architecture that enables robots to perform complex, long-horizon tasks in real-world environments with improved reasoning, control, and adaptability.

• What is the ER↔VLA stack?

  The ER↔VLA stack is a two-model architecture at the core of Gemini Robotics 1.5. It separates embodied intelligence into Gemini Robotics-ER 1.5 (for high-level reasoning and planning) and Gemini Robotics 1.5 (for low-level visuomotor control).

• What is Gemini Robotics-ER 1.5 responsible for?

  Gemini Robotics-ER 1.5 is the “reasoner” or “orchestrator.” It’s a multimodal planner that handles high-level tasks like scene understanding, dynamic planning, tracking task progress, estimating success, and invoking external tools (e.g., web search, APIs) to inform its decisions.

• What is Gemini Robotics 1.5 (VLA) responsible for?

  Gemini Robotics 1.5 (VLA) acts as the low-level Vision-Language-Action controller. It translates instructions and sensory input into precise motor commands and generates “think-before-act” traces to break down complex tasks into manageable sub-skills, aiding in error recovery and robust execution.

• What is Motion Transfer and why is it important?

  Motion Transfer (MT) is a capability that allows skills learned on one robot to be transferred zero-shot to entirely different, heterogeneous robot platforms. This is crucial because it vastly reduces the need for extensive, robot-specific data collection and retraining, accelerating deployment and reducing development costs.

• How does Gemini Robotics 1.5 ensure safety?

  DeepMind prioritizes safety through layered controls, including policy-aligned dialog, safety-aware grounding (to prevent hazardous actions), and adherence to physical limits. They also use advanced evaluation techniques like ASIMOV-style scenario testing and auto red-teaming to identify and mitigate edge-case failures and potential unsafe behaviors before physical deployment.

• How can developers access Gemini Robotics 1.5?

  Developers can currently access Gemini Robotics-ER 1.5 via the Gemini API in Google AI Studio. The Gemini Robotics 1.5 VLA controller is initially available to selected partners as it rolls out.