Gemini Robotics 2: Whole-Body Control for Humanoids

Introduction

Google DeepMind announced Gemini Robotics 2 on July 30, 2026, describing it as the intelligence layer for its next generation of robots. The release moves the company's robotics stack beyond tabletop manipulation, the primary focus of earlier Gemini Robotics releases, into whole-body control, five-finger dexterity, and coordination between multiple robots working on the same task. Rather than a single model, Gemini Robotics 2 ships as three separate models designed to work together: a vision-language-action (VLA) model for physical control, an embodied-reasoning model for high-level planning, and an efficient on-device model for local execution without a network connection.

Feature Overview

Three Models, Three Roles

Gemini Robotics ER 2 operates with a 128,000-token context window and can emit up to 64,000 tokens, allowing it to reason over long, multi-modal task histories before issuing instructions to the VLA layer. Gemini Robotics On-Device 2 takes text, images, and robot proprioception (the robot's own sense of its joint positions and movement) as numeric input and outputs robot actions directly as numeric values, without needing a live connection to the cloud.

Whole-Body Control and Cross-Embodiment

DeepMind demonstrated whole-body control on Apptronik's Apollo 2 humanoid, showing multi-step tasks such as placing a watering can into a green bin on a bottom shelf, tasks that require walking, grasping, and precise placement in sequence rather than manipulation from a fixed stance. DeepMind acknowledges in its own materials that the robots "have more to advance in movement speed," an explicit limitation rather than a claim of finished performance.

A notable architectural claim is cross-embodiment generalization: a single Gemini Robotics 2 model checkpoint operates across three distinct hardware configurations, Apollo 2 fitted with SharpaWave hands, Apollo 2 fitted with Inspire hands, and a Franka Duo arm fitted with a Robotiq gripper, without retraining a separate model per robot.

Multi-Robot Collaboration

The release also introduces coordination between different robot types. In a demonstration pairing Apptronik's Apollo 2 with a Franka F3 Duo arm, the two robots communicated through a shared semantic understanding of the task to hand off subtasks to each other, rather than each operating independently on its own portion of a job.

Usability Analysis

DeepMind published granular benchmark numbers alongside the release rather than a single aggregate score, which gives a more concrete picture of where the system is strong and where it isn't.

Dexterity remains the system's weakest axis by DeepMind's own characterization, with task-level success rates ranging from 32% on dustpan handling up to 92% on unscrewing a light bulb. That spread illustrates that fine manipulation performance depends heavily on the specific task's geometry and grip requirements, rather than dexterity being uniformly reliable or uniformly weak.

On the reasoning side, DeepMind reports that ER 2 executes roughly 4 times faster than larger competing models while matching their accuracy on moment-finding, attributing part of the speed gain to the Gemini Live API's bidirectional streaming, which removes the "stop-and-think pauses that break multi-step execution" found in earlier designs.

Gemini Robotics On-Device 2 is built for scenarios without reliable connectivity and is designed to adapt to new bi-arm robot embodiments quickly, typically within a few hours using fewer than 200 examples. DeepMind's own test on the previously unseen SO101 platform showed performance rising from 6.7% success with On-Device 1 to 53.3% with On-Device 2, a concrete measure of how much the on-device model's few-shot adaptation improved between generations.

Access

Availability is tiered and, for most developers, limited. Gemini Robotics ER 2 is in public preview through the Gemini API and Google AI Studio under the model string gemini-robotics-er-2-preview, and in private preview on the Gemini Enterprise Agent Platform. The core VLA model is restricted to early-access partners. Gemini Robotics On-Device 2 is limited to participants in Google's Trusted Testers program. Sample code is published in DeepMind's robotics-samples GitHub repository. Hardware partners named alongside the release include Apptronik, Franka, Boston Dynamics (Spot), Dexmate, SO101, and Trossen.

Pros and Cons

Pros

• A single Gemini Robotics 2 checkpoint generalizes across three distinct hardware configurations, reducing the need to retrain a separate model per robot.
• Gemini Robotics ER 2's roughly 4x execution speed and bidirectional streaming remove stop-and-think pauses that previously interrupted multi-step task execution.
• On-Device 2's few-shot adaptation, fewer than 200 examples to reach usable performance on a new embodiment, is measurably better than its predecessor, as shown by the SO101 jump from 6.7% to 53.3%.
• Multi-robot task handoff between different robot types (Apollo 2 and Franka F3 Duo) demonstrates coordination beyond single-robot manipulation.
• DeepMind published granular, task-level benchmark numbers rather than a single marketing score, giving outside evaluators a transparent baseline.

Cons

• Dexterity success rates vary widely by task, from 32% to 92%, and DeepMind itself identifies dexterity as the system's weak axis.
• Whole-body manipulation accuracy drops as the target location gets lower, from 76.3% at shelf height down to 45.7% on the floor, indicating harder lower-workspace tasks are less reliable.
• Access to the core VLA model and On-Device 2 is restricted to early-access partners and Trusted Testers respectively, so most developers can only experiment with the ER 2 preview.
• DeepMind explicitly notes robot movement speed still needs improvement, an acknowledged gap for real-world deployment timelines.

Outlook

Gemini Robotics 2 represents an incremental but concrete step from tabletop-only manipulation toward whole-body, cross-embodiment robot control, with published numbers that make the current gaps, particularly in dexterity and floor-level manipulation, visible rather than obscured. The breadth of named hardware partners, spanning humanoids (Apptronik), industrial arms (Franka), and quadrupeds (Boston Dynamics Spot), suggests Google is positioning the Gemini Robotics line as a general-purpose intelligence layer that multiple robot manufacturers can adopt, rather than software tied to a single hardware platform.

The release also arrived the same week Google DeepMind reassigned staff from its AlphaFold team to Gemini-focused projects and Isomorphic Labs, an internal signal that the company is shifting research resources toward applied, product-facing AI work, including physical AI, alongside its longer-running scientific research efforts.

Conclusion

Gemini Robotics 2's three-model release, spanning a VLA for physical control, an embodied-reasoning planner, and an efficient on-device model, extends Google DeepMind's robotics stack from tabletop tasks into whole-body and multi-robot scenarios, backed by detailed, mixed benchmark results rather than a single success metric. The system is not yet uniformly dexterous or fast, and broad access remains limited to preview and partner programs for now. It is most relevant to robotics researchers and hardware partners already integrated with Google's ecosystem, and less immediately useful to independent developers who cannot yet access the VLA or On-Device 2 models outside of partner and Trusted Tester programs.

Editor's Verdict

Gemini Robotics 2: Whole-Body Control for Humanoids earns a solid recommendation within the gemini space.

The strongest case for paying attention is A single VLA checkpoint generalizes across three distinct hardware configurations, reducing the need to retrain a model per robot, which raises the bar for what readers should now expect from peers in this space. Reinforcing that, ER 2's roughly 4x execution speed and bidirectional streaming remove stop-and-think pauses in multi-step task execution adds practical value rather than just headline appeal. The broader signal worth registering is straightforward: A single Gemini Robotics 2 (VLA) checkpoint generalizes across three distinct hardware setups, Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and Franka Duo with a Robotiq gripper, without per-robot retraining. On the other side of the ledger, dexterity success rates vary widely by task (32% to 92%), and DeepMind itself calls dexterity the system's weak axis is a real constraint, not a marketing footnote, and it should factor into any serious decision. Layered on top of that, whole-body manipulation accuracy falls sharply for lower-workspace tasks, from 76.3% at shelf height to 45.7% on the floor narrows the set of teams for whom this is an obvious yes.

For Google Cloud and Workspace integrators, multimodal-first teams, and Gemini API adopters, this is a serious evaluation candidate, not just a curiosity to bookmark. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.