Gemini Robotics 2: The Planner-Executor Pattern Goes Physical

Google DeepMind announced Gemini Robotics 2 today, and the headline — “whole body intelligence for robots” — undersells what’s actually interesting about the release. Yes, the model can now control an entire humanoid from feet to fingertips. Yes, it can tie knots with a 22-degree-of-freedom hand. But the thing that made me stop and read the whole announcement twice is the architecture: DeepMind has settled on the same planner-executor split that those of us building multi-agent software systems converged on, and they’ve added a safety layer whose job is literally to refuse unsafe tool calls.

If you build agent pipelines, this release is a mirror. It’s worth looking into.

The Three-Model Architecture

Gemini Robotics 2 isn’t one model. It’s three, with clearly separated responsibilities:

Gemini Robotics 2 is the vision-language-action model (VLA). It converts vision and language input directly into motor control. This is the executor — the model that actually moves the robot. It controls full humanoids (the demos run on Apptronik’s Apollo 2) and bi-arm platforms, and it handles dexterous manipulation on both five-fingered hands and standard parallel grippers.

Gemini Robotics ER 2 is the embodied reasoning model — a vision-language model that acts as the agent. It doesn’t move anything. It observes the environment, reasons about the steps needed to complete a task, coordinates with the VLA to execute actions, and tracks progress until the task is done. DeepMind describes it as the robot’s “high-level brain.” ER 2 handles multi-step tasks lasting several minutes and involving hundreds of individual decisions, and it now understands when tasks begin and end, pinpointing the moment key events occur.

Gemini Robotics On-Device 2 is the efficiency play: a VLA optimized to run locally on the robot itself, for deployments where network latency or connectivity constraints rule out a cloud round-trip. It’s natively multi-embodiment and inherits the motion-transfer techniques from Gemini Robotics 1.5 — adaptation to a completely new bi-arm robot body takes a few hours, typically with fewer than 200 examples.

ER 2 is available today on Google AI Studio and in private preview on the Gemini Enterprise Agent Platform. The VLA and On-Device models are gated to early-access partners.

Why the ER/VLA Split Matters

Here’s the part that feels familiar. In the multi-agent systems I work on daily, the single most reliable architectural pattern is the separation of reasoning from execution. A planner model — given a goal and a view of the world — decomposes the task into steps, decides what to do next, and evaluates whether the last action succeeded. An executor takes a single well-specified step and carries it out. The planner never writes files; the executor never re-plans.

DeepMind has built exactly this, for physics. ER 2 is the planner. The VLA is the executor. The interface between them is a stream of action instructions instead of tool calls, but the information flow is identical: high-level intent flows down, execution results flow back up, and the reasoning layer self-corrects when a step fails.

This convergence isn’t a coincidence, and it isn’t DeepMind copying software agents — the ER/VLA split dates back to the original Gemini Robotics in March 2025 and RT-2 before it. It’s the same answer to the same constraint. A model that’s good at deliberate, multi-step reasoning over minutes-long horizons has different training data, different context structure, and different latency tolerance than a model that has to emit motor commands at control-loop frequency. Cramming both into one model means compromising both. Splitting them means each can be optimized, evaluated, and improved independently.

There’s a second benefit that anyone who’s run a staged agent pipeline will recognize: the split gives you a place to put the safety checks. Which brings me to the most interesting part of the announcement.

ASIMOV-Agentic: The Safety Layer That Refuses Tool Calls

Buried in the safety section is a detail that deserves more attention than it will get. DeepMind is introducing ASIMOV-Agentic, a benchmark for agentic safety orchestration. Two of the things it measures:

1. The embodied reasoning agent’s ability to refuse unsafe tool calls from the VLA.
2. The agent’s ability to predict whether a task is possible and proactively request human intervention when uncertain.

Read that first one again. The reasoning model sits between the user and the action model, and one of its measured competencies is overriding the executor when the executor proposes something unsafe. This is the paranoid-reviewer pattern — a stage in the pipeline whose job is not to do the work but to evaluate whether the work should happen at all.

In software agent systems, this pattern shows up as a review gate: one agent produces a change, another agent with a different prompt and a different objective — adversarial review, security scan, policy check — decides whether it ships. The reviewer is dumber than the producer in general capability but sharper along one specific axis, and its veto is structural, not advisory. ASIMOV-Agentic is measuring whether ER 2 can play that role for a robot: the VLA proposes an action sequence, the reasoning layer evaluates it against safety constraints, and unsafe proposals get refused before any motor command executes.

The second capability — calibrated uncertainty — is the one I find hardest in software agents and it’s encouraging to see it measured explicitly here. Knowing when a task is beyond you, and escalating to a human before attempting it rather than after failing, is the difference between an agent you can supervise and an agent you have to watch. DeepMind says ER 2 is their safest robotics model to date on safety-constraint-following and human-proximity benchmarks — it detects nearby humans, triggers safety tool calls, and brings the robot to a stop if someone gets too close. There’s a full safety technical report alongside the release.

The Embodiment Problem Is the Tool-Adaptation Problem

The On-Device 2 claims deserve their own scrutiny, because they address the robotics equivalent of a problem every agent builder knows: adapting a general model to a new interface.

A software agent that works beautifully against one API often falls apart against another with different conventions, different error formats, different rate limits. The robotics version is embodiment transfer — a policy trained on one robot body, with its specific geometry, sensors, and degrees of freedom, traditionally fails on a different body. Retraining for each new platform is what made general-purpose robotics uneconomical.

DeepMind’s claim: On-Device 2 adapts to a completely new bi-arm embodiment in a few hours with fewer than 200 examples, even for bodies with drastically different shapes, sensors, and DoF counts. The demos show the same checkpoint driving the Dexmate, SO101, and Trossen platforms. The main Gemini Robotics 2 VLA also demonstrates single-checkpoint control across Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and a Franka Duo with a Robotiq gripper.

Fewer than 200 examples for a new body. In software terms, that’s few-shot adaptation to a new tool schema — and it implies the model has learned something like the concept of manipulation rather than the specifics of one manipulator. If the claim holds up outside partner demos, the economics of deploying robots across heterogeneous fleets change substantially. The per-platform engineering tax is what keeps most robotics deployments locked to a single vendor’s hardware.

Multi-Robot Collaboration

One more capability worth flagging: ER 2 introduces multi-robot collaboration — different types of robots communicating and dividing labor on workflows no single robot could complete alone. The example given is robots teaming up to clean a cluttered room faster.

This is early, and the announcement is thin on mechanism. But the direction matters: the reasoning layer isn’t just orchestrating one body’s actions, it’s becoming an orchestrator across agents, each with its own executor. That’s the same hierarchical shape — orchestrator, worker pool, per-worker execution model — that scales software agent systems past what a single agent loop can handle.

The Honest Limitations

Credit to DeepMind for publishing the success-rate chart with the release, because it tells a more tempered story than the demo videos. Whole-body tasks and gripper-based dexterous tasks land at medium-to-high success rates. Multi-finger dexterous manipulation — the knot-tying, ziplock-sealing category — remains genuinely hard, with individual task performance shown precisely because the averages aren’t flattering. Movement speed is called out explicitly: the robots have “more to advance in movement speed,” which is a polite way of saying the Apollo 2 demos walk slowly.

Access is also narrow. ER 2 on AI Studio is the only piece most developers can touch today; the models that actually move hardware are partner-gated. And multi-minute, hundreds-of-decisions task horizons sound impressive until you price the failure compounding: at 95% per-step reliability, a 200-step task succeeds about 0.0035% of the time. The self-correction loop is what rescues this — the reasoning layer catching and retrying failed steps — which is exactly why the ER/VLA split with progress tracking is load-bearing rather than cosmetic.

Why This Release Sticks With Me

Most robotics announcements are about hardware or about a single impressive demo. This one is about an architecture, and the architecture is the argument: general-purpose physical intelligence decomposes the same way general-purpose digital intelligence does. A reasoner that plans, tracks progress, and knows when to ask for help. An executor that turns instructions into actions. A safety layer with structural veto power. An on-device variant for when the network can’t be in the loop. Fast adaptation to new bodies the way good agents adapt to new tools.

The models will keep improving. The knot-tying success rates will climb, the walking will speed up, the embodiment transfer will get cheaper. But the shape of the system — planner, executor, gate — is the part I’d bet on surviving every future revision, because it’s the shape that works everywhere else we’ve tried to build machines that act on intent.

Gemini Robotics ER 2 is in Google AI Studio today if you want to probe the reasoning layer yourself. The model cards and the safety technical report are linked from DeepMind’s announcement. The rest of us get to watch what the early-access partners build with the hands.