"Gemini Robotics: When AI Finally Learns to Touch the Physical World"

For two years, the AI industry has competed on benchmarks that measure how well models write code, summarize documents, and pass multiple-choice exams. Gemini Robotics shifts the competition to a different arena: the physical world.

Google DeepMind's Gemini Robotics is a family of models built on Gemini 2.0 and 3.0 foundations that can perceive visual scenes, reason about spatial relationships, and output motor commands that drive actual robots. Carolina Parada, Senior Director at DeepMind, frames the ambition plainly: "We developed our Gemini Robotics models to bring AI into the physical world."

The significance is structural. Every major AI lab has recognized that language-only intelligence hits a ceiling. Understanding physics, manipulating objects, navigating unstructured environments: these capabilities require an AI that can act, not just talk. Gemini Robotics is Google's answer, and the benchmark numbers suggest it is currently ahead of the pack by a wide margin.

The Three-Model Family

Gemini Robotics is not a single model. It is a coordinated family with distinct roles, each addressing a different bottleneck in physical AI.

Gemini Robotics 1.5 (VLA) is the action model. VLA stands for Vision-Language-Action. It takes visual input and natural-language instructions, then directly generates motor commands for a robot to execute. This is the model that actually moves things. It represents a core architectural bet: a large enough language model already contains an implicit world model, so no separate physics simulator is needed to produce useful physical behavior.

Gemini Robotics-ER 1.6 (Embodied Reasoning) is the cognitive layer. ER handles the tasks that require thinking before acting: reading instrument panels, identifying which object to pick up, planning a sequence of grasps, and determining whether a previous action succeeded. It excels at spatial reasoning and affordance detection, meaning it can look at a scene and understand what actions are physically possible.

Gemini Robotics On-Device is the deployment variant. It compresses the VLA architecture to run locally on robot hardware, reducing latency to the point where real-time control loops become feasible. This is essential for any production scenario where cloud round-trips introduce unacceptable delay. A factory robot that pauses for two seconds while waiting for a server response is a factory robot that stops the assembly line.

Together, the three models cover the full pipeline: reason about what to do (ER), decide how to do it (VLA with optional Thinking Mode), and execute it fast enough to matter (On-Device). This separation of concerns is itself worth noting. Rather than building a single monolithic model that tries to do everything, DeepMind has decomposed the problem into perception-reasoning, action-planning, and low-latency execution. Each component can be improved independently, and each can be deployed on the hardware best suited to its compute requirements.

Technical Architecture: How It Works

VLA: From Pixels to Motor Commands

The VLA architecture is the core innovation. Traditional robotics pipelines split perception, planning, and control into separate modules, each hand-engineered. VLA collapses them into a single end-to-end model. Visual tokens from a camera feed and text tokens from a language instruction are processed together, and the output is a stream of continuous motor commands: joint angles, velocities, or end-effector poses.

This is possible because the underlying Gemini foundation model has already learned rich representations of objects, spatial relationships, and physical concepts from its training on text and images. The VLA fine-tuning layer teaches the model to map those representations onto actionable outputs.

Embodied Reasoning: Spatial Intelligence

ER 1.6 extends the Gemini model with capabilities specific to physical interaction. It can identify objects in cluttered scenes, estimate 3D spatial relationships from 2D images, and evaluate whether a planned action is physically feasible. Crucially, it also performs success detection: after a robot attempts a task, ER can judge whether it succeeded using either a single camera view (86% accuracy) or multiple views (93% accuracy).

This matters because real-world robotics fails constantly. Grips slip, objects shift, lighting changes. A robot that can detect its own failures and retry is qualitatively different from one that blindly continues executing a failed plan. The 93% multi-view success detection rate suggests that ER 1.6 is approaching the reliability threshold needed for autonomous operation in supervised settings, though unsupervised deployment will likely require further improvement.

Thinking Mode

One of the more interesting architectural choices is Thinking Mode, which applies to the VLA model. Before generating motor commands, the model produces a verbal chain-of-thought describing its reasoning. For example: "The user wants me to put the apple in the red bowl. The apple is on the left side of the table. The red bowl is behind the cup. I should grasp the apple from above, then move it to the bowl."

This is analogous to the thinking tokens in o1 and o3-style language models, but applied to physical planning. Early results suggest it improves performance on multi-step tasks where naive action generation fails.

Motion Transfer

A persistent problem in robot learning is that data from one robot platform does not transfer cleanly to another. Different robots have different joint configurations, end-effectors, and kinematic structures. Motion Transfer is DeepMind's solution: a technique that unifies data from heterogeneous robot platforms into a single shared representation, allowing a single model to control robots with very different bodies.

This is a prerequisite for any general-purpose robotics model. Without it, you would need to train a separate model for every robot hardware variant.

The Benchmark Numbers

DeepMind has published detailed benchmarks for ER 1.6, and the results are striking. The model does not just edge out competitors. It often more than doubles their performance on generalization tasks.

Instrument Reading

The instrument reading benchmark tests whether a model can interpret analog gauges, displays, and measurement devices from images. This is a practical skill for industrial robotics.

The jump from ER 1.5 (23%) to ER 1.6 (86%) is a 3.7x improvement in a single generation. Adding agentic vision tool use pushes it to 93%, suggesting that the model benefits from iterative visual examination the way a human would lean in to read a small dial.

Point-Bench: Spatial Reasoning

Point-Bench measures a model's ability to understand spatial relationships in images: where objects are, how many there are, what actions they afford. The numbers below compare ER 1.5 against the strongest language model competitors.

ER 1.5 scores 52.6 on overall average. GPT-5 manages 30.8. On spatial reasoning specifically, ER 1.5 nearly doubles GPT-5's score (61.7 vs 33.0). The counting benchmark shows an even wider gap: 86.8 vs 53.7.

These are the tasks that matter most for a robot trying to operate in a kitchen, warehouse, or construction site. A model that cannot count objects or understand affordances will fail at the most basic manipulation tasks.

Generalization Performance

Across the full suite of generalization benchmarks, DeepMind reports that Gemini Robotics more than doubles performance compared to other state-of-the-art VLA models. This is a single summary statistic, but the magnitude is unusual. In most AI benchmark races, the leader edges out competitors by 5-15%. A 2x advantage suggests a genuine architectural lead, not just incremental tuning.

The Competition: Three Philosophies of Physical AI

The VLA model space is heating up rapidly. Three major players have articulated distinct philosophies about how to build AI that can interact with the physical world.

NVIDIA GR00T N1.7: Physics First

NVIDIA's approach starts from the physics. GR00T N1.7 trains a world model (Cosmos) first, building an explicit understanding of physical dynamics, then layers a VLA on top. The architecture uses a dual system: System 2 handles slow, deliberative reasoning, while System 1 generates fast reflexive actions. NVIDIA has invested heavily in data infrastructure, compiling EgoScale, a dataset of over 20,000 hours of first-person video.

The philosophical commitment here is that physical intelligence requires an explicit physics engine in the loop. The risk is that building a world model that generalizes across all physical scenarios is an enormously hard problem, and errors in the world model cascade into the action model.

Google Gemini Robotics 1.5: Implicit World Knowledge

Google's bet is the opposite: a large enough language model already contains an implicit world model, learned from the vast corpus of text and image data it was trained on. There is no separate world model component. The VLA fine-tuning layer simply teaches the model to map its existing understanding onto motor commands, and Thinking Mode provides explicit reasoning when needed.

The advantage is architectural simplicity and leverage: every improvement to the base Gemini model automatically improves the robotics model. The risk is that implicit physical understanding has gaps that an explicit physics model would catch.

Physical Intelligence pi0.5: Direct Action Learning

Physical Intelligence, the company behind pi0.5, argues that direct action data learning is the most efficient path. Their model uses Flow Matching on a PaliGemma 3B base and has been trained on roughly 400 hours of real household robot data. The approach is lean and focused: skip the world model, skip the general knowledge, and learn directly from demonstrations of the tasks you care about.

This is pragmatic and works well in constrained domains. The open question is whether it generalizes as broadly as the other two approaches when confronted with novel environments and tasks.

Comparison at a Glance

These are early days. None of these approaches has been validated at industrial scale yet. But the philosophical divergence matters because it determines where each system will break first: NVIDIA at the world-model boundary, Google at the implicit-knowledge gap, and PI at the generalization horizon.

Hardware Partners: The Bodies Behind the Brains

A model without a body is a simulation. DeepMind has lined up two major hardware partners to put Gemini Robotics into physical form.

Boston Dynamics and the New Atlas

In January 2026, at CES, Boston Dynamics and Google DeepMind announced a formal partnership to integrate Gemini Robotics with the new all-electric Atlas humanoid robot. This is a significant alignment of capabilities. Boston Dynamics has long been recognized for the athletic performance of its robots: backflips, parkour, dynamic balance. But the company only announced its intention to build a commercial humanoid in 2024, after it became clear that recent AI advances had accelerated the pace of how robots could be trained and deployed.

The partnership combines DeepMind's software intelligence with Boston Dynamics' unmatched hardware engineering. Atlas provides the physical platform: powerful actuators, sophisticated balance control, and robust sensing. Gemini Robotics provides the brain that can interpret scenes, plan actions, and adapt to novel situations.

This is a notable reunion. Google (now Alphabet) owned Boston Dynamics from 2013 to 2017 before selling it to SoftBank, which later sold it to Hyundai. The companies share institutional DNA. The current partnership is structured as a technology integration rather than an acquisition, which gives both sides flexibility while still combining what are arguably the strongest software and hardware capabilities in the robotics industry.

Apptronik and Apollo

Apptronik, announced as a partner in December 2024, brings the Apollo humanoid to the table. Apollo is a more compact platform designed for logistics and manufacturing environments. The Gemini Robotics integration with Apollo focuses on tasks like bin picking, assembly assistance, and warehouse navigation.

Why the Partnerships Matter

DeepMind's framing is telling: "Gemini Robotics models allow robots of any shape and size to perceive, reason, use tools and interact with humans." The phrase "any shape and size" is a direct reference to Motion Transfer, the technique that allows a single model to control different robot platforms. If it works as advertised, DeepMind becomes the default AI layer for the entire robotics industry, independent of hardware vendor.

The Genie 3 Connection: From Simulation to Reality

Gemini Robotics does not exist in isolation. It sits at the end of a technology progression within DeepMind that connects world simulation to physical execution, and understanding this progression reveals the larger strategy.

Both Genie 3 and Gemini Robotics fall under the "World models and embodied AI" research group at DeepMind. Genie 3, which we examined in our earlier analysis of DeepMind's world simulation work, is a world simulator: a model that can generate interactive virtual environments from images or text descriptions. It generates the training ground.

The progression looks like this: Genie 3 creates realistic, interactive virtual environments. SIMA (Scalable Instructable Multiworld Agent) learns to act within those virtual environments, developing generalizable skills across game worlds and simulations. Gemini Robotics takes the same principles and applies them to the physical world, with real sensors and real actuators.

DeepMind describes Genie 3 as "a key stepping stone on the path to AGI, enabling AI agents capable of reasoning, problem solving, and real-world actions." The wording is deliberate. The path goes through world simulation, then virtual agents, then physical agents. Each stage builds on the representations learned in the previous one.

This also explains the implicit world model bet in Gemini Robotics. If Genie 3 can simulate physical worlds accurately enough, you can generate unlimited training data for robotics models without ever touching a physical robot. The world model does not need to be embedded in the robotics model itself. It can live upstream, in the data generation pipeline.

The implications extend beyond training efficiency. A world simulator like Genie 3 can also be used for safety testing: running a robot policy through millions of simulated scenarios before ever deploying it on hardware. It can generate adversarial edge cases that human engineers would never think to test. And it can provide the "experience" base that a purely data-driven VLA model needs to generalize beyond its direct training distribution. In this sense, Genie 3 is not just a research curiosity. It is the infrastructure layer that makes the implicit world model approach viable at scale.

What This Means: From Code Generation to Physical Interaction

The dominant narrative in AI over the past three years has been about productivity: models that write code faster, summarize meetings, generate marketing copy. Gemini Robotics represents a qualitative shift. It graduates from software to physical manipulation, from generating text to moving objects.

Three implications stand out.

First, the generalization numbers matter more than the peak performance numbers. A robot that achieves 95% accuracy on a single task in a controlled lab is less valuable than one that achieves 70% across a wide range of novel tasks in unstructured environments. The 2x generalization advantage of Gemini Robotics over competing VLA models is the metric to watch.

Second, the Boston Dynamics partnership is a signal. Boston Dynamics spent decades building robots with hand-engineered control systems. Their decision to adopt an AI-first approach with Gemini Robotics indicates that the industry has crossed a threshold where learned controllers are outperforming engineered ones, at least for high-level planning and perception.

Third, the competition between the three VLA philosophies will be resolved by deployment data, not by benchmarks. NVIDIA's physics-first approach, Google's implicit-knowledge approach, and Physical Intelligence's direct-learning approach all have plausible theoretical justifications. The market will decide based on which one actually works in production: in warehouses, factories, hospitals, and homes.

There is also a fourth player worth watching: Figure AI, whose Helix system takes yet another angle on the problem by training separate vision-language and control policies that communicate through a shared latent space. The field is still early enough that the winning architecture has not been settled, and it is possible that different approaches will dominate different verticals. NVIDIA's explicit world model may prove superior for high-precision manufacturing, while Google's implicit approach may win in unstructured environments like homes and hospitals where variability is the defining challenge.

DeepMind's stated mission is "powering an era of physical agents to transform how robots actively understand their environments." If the benchmarks hold up in real-world deployment, and if the Motion Transfer technology truly allows a single model to control robots of any shape and size, then the robotics industry is about to undergo the same platform consolidation that the software industry experienced when foundation models emerged. The question is no longer whether AI can reach the physical world. It has. The question now is how fast it scales.

Sources: Google DeepMind Gemini Robotics | Gemini Robotics Blog Post | Gemini Robotics-ER 1.6 Blog Post | ArXiv: Gemini Robotics (2503.20020) | ArXiv: ER 1.6 (2510.03342) | Boston Dynamics Partnership | VLA Architecture Comparison (Pebblous)