Google DeepMind launches Gemini Robotics 2 and Gemini Robotics-ER 2 for whole-body robot control

Google DeepMind introduced Gemini Robotics 2 on July 30, 2026, describing it as "the intelligence layer powering the next generation of truly adaptable robots." The release pairs a new vision-language-action (VLA) model with an upgraded reasoning model, Gemini Robotics-ER 2, aimed at giving robots the physical dexterity and situational judgment to handle open-ended, real-world tasks rather than narrow, pre-scripted ones.

What's new

The headline capability is whole-body control: Gemini Robotics 2 "unlocks intelligent whole-body control, advanced dexterity, and multi-robot collaboration," enabling a humanoid to "walk, crouch, stretch, and manipulate objects to clean up a cluttered room." That marks a step up from earlier robotics models that mostly focused on isolated arm or gripper manipulation.

Alongside it, Gemini Robotics-ER 2 acts as what DeepMind calls "a high-level brain for robots," handling video understanding, task orchestration, and multi-robot coordination. It "allows robots to chat with humans, understand the physical world, and plan multi-step tasks," and can "watch continuous video feeds" to "track their own progress, adapt if something goes wrong, and know exactly when to move on to the next step."

Availability is split by maturity. Gemini Robotics-ER 2 is "now publicly available to developers via the Gemini API, Google AI Studio, and in private preview on Gemini Enterprise Agent Platform." The VLA and on-device models, by contrast, are limited to early-access partners, with DeepMind inviting interested developers to "sign up for our Trusted Tester Program."

Context

The release lands alongside a broader push into embodied AI across the industry — NVIDIA, Hugging Face, and others have all shipped robotics-focused models and tooling in recent weeks. Google's approach leans on its existing Gemini model family: rather than building a separate robotics stack from scratch, DeepMind is extending Gemini's multimodal reasoning into physical control, with the ER 2 reasoning layer reusing the same API and Studio distribution channels developers already use for Gemini's chat and agent products.

The reasoning/action split — a fast, publicly available perception-and-planning model paired with a more restricted action model — mirrors a pattern seen elsewhere in robotics AI, where labs are more cautious about broadly releasing models that directly actuate physical hardware than ones that only interpret and plan.

Why it matters

Whole-body control is the harder half of the humanoid-robotics problem: coordinating balance, locomotion, and manipulation simultaneously, rather than treating them as separate subsystems. If Gemini Robotics 2 delivers on multi-robot collaboration and adaptive task recovery, it pushes Google further into direct competition with robotics-focused labs and hardware makers building toward general-purpose humanoids for warehouses, homes, and service settings.

The gated rollout is also a signal in itself. By making the reasoning model (ER 2) broadly available while keeping the action model restricted to trusted partners, Google is separating the lower-risk "understand and plan" capability from the higher-risk "directly control a physical robot" capability — a distinction likely to shape how other labs stage their own robotics releases as the category matures.