End-to-End Humanoid Control Arrives in Gemini Robotics 2

On July 30, 2026, Google DeepMind released the Gemini Robotics 2 series, transitioning its physical AI stack from upper-body tabletop manipulation to end-to-end whole-body control. The release introduces a suite of three models designed to coordinate locomotion, torso stability, and fine dexterity for humanoid and bi-manual robots. If you build robotics software, this shifts the boundary of what can be handled by a single vision-language-action model versus deterministic control systems.

The Three-Model Architecture

The Gemini Robotics 2 release divides robotic operation into distinct layers, separating high-level planning from localized motor control.

The flagship Gemini Robotics 2 (VLA) serves as the primary motor control layer. It translates visual and language inputs directly into physical movements. Unlike the 1.5 generation, this model supports simultaneous whole-body motions, allowing a humanoid robot to walk, crouch, and reach for an object in a single fluid action sequence.

For task orchestration and long-horizon planning, DeepMind introduced Gemini Robotics ER 2 (VLM). This Embodied Reasoning model acts as the high-level brain for robotic fleets. Based on Gemini 3.5 Flash, the ER 2 model uses a 128k context window to process interleaved video, audio, and text. This model is specifically designed for multi-robot collaboration, allowing you to orchestrate multi-agent systems where diverse hardware shares a semantic understanding of an environment.

The final layer is Gemini Robotics On-Device 2. Based on the Gemma architecture, this lightweight VLA is built for local execution on robotic hardware. DeepMind reports that this on-device model can achieve embodiment transfer to entirely new robotic body types in a few hours using fewer than 200 training examples.

Capabilities and Task Precision

DeepMind demonstrated these capabilities using Apptronik’s Apollo 2 humanoid robot equipped with SharpaWave 22-degree-of-freedom hands. In end-to-end demonstrations, the models successfully directed the robot to autonomously navigate a room, retrieve a specific object, walk to a shelving unit, and place it precisely on a designated shelf.

The models show high success rates for extraction and placement, though complex re-insertion tasks remain a constraint.

Hardware Task	Success Rate
Unscrewing a lightbulb	92.0%
Precise two-finger insertion	89.6%
Screwing a lightbulb back in	36.0%

To address the safety implications of autonomous hardware, DeepMind released the ASIMOV-Agentic Benchmark on Hugging Face. When evaluating agents for physical deployment, this benchmark measures a model’s ability to refuse unsafe tool calls, recognize impossible tasks, and request human intervention when environmental uncertainty is high.

API Integration and Availability

Gemini Robotics ER 2 is available in public preview. Developers can access it via Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform.

Google introduced two distinct API endpoints for this release. The standard gemini-robotics-er-2-preview endpoint handles standard inference, while gemini-robotics-er-2-streaming-preview is optimized for the Gemini Live API to deliver the low-latency responses required for real-time robotic adjustments.

The Gemini Robotics 2 VLA and On-Device 2 models remain restricted to early-access hardware partners, including Apptronik, Boston Dynamics, and Agile Robots.

If you are integrating the ER 2 model into an existing robotics stack, you will need to map its high-level orchestration outputs to your current motor control software until the VLA models reach general availability.