Gemini Robotics 2: What the Whole-Body, Reasoning, and On-Device Models Actually Prove

The short answer

Gemini Robotics 2 is a three-model stack, not one robot product. Google DeepMind separates low-level action, high-level embodied reasoning, and local inference:

1. Gemini Robotics 2 converts vision and language into robot actions. Google demonstrates it controlling full-body movement and manipulation on Apptronik Apollo 2, plus bi-arm manipulation on Franka Duo.
2. Gemini Robotics ER 2 watches continuous video, plans multi-step work, calls control tools, tracks progress, and coordinates robots. It outputs text or tool instructions rather than joint actions.
3. Gemini Robotics On-Device 2 runs locally and can be adapted to a new bi-arm embodiment with a few hours of data, typically fewer than 200 examples according to Google.

The strongest evidence is cross-embodiment first-party demonstration, public access to the reasoning layer, explicit model cards, and some disclosed reasoning benchmarks. The largest gaps are complete task-level denominators for the VLA, intervention and reset rules, motion speed, independent replication, and recurring customer operation.

This means the launch is important model evidence. It is not evidence that an Apollo 2 powered by Gemini Robotics 2 is commercially deployed or ready for unsupervised safety-critical work.

Three models, three jobs, three access states

Model	Primary job	Input and output	Demonstrated hardware	Public access on July 31, 2026	Evidence boundary
Gemini Robotics 2	Whole-body and dexterous motor control	Images, language, robot state to numerical actions	Apollo 2 with two hand configurations, Franka Duo	Early-access partners	First-party demos and aggregated skill results, without complete public trial logs
Gemini Robotics ER 2	Planning, progress tracking, tool orchestration, and robot collaboration	Text, image, video, and audio to text or tool calls	Spot, Apollo 2, Franka Duo, simulation, VLA, and human teleoperation setups	Gemini API and Google AI Studio, plus private enterprise preview	Public software access and disclosed benchmarks, but not a complete autonomous robot stack
Gemini Robotics On-Device 2	Local action inference and rapid adaptation	Images, language, and proprioception to numerical actions	Dexmate, SO-101, Trossen, and other bi-arm setups	Selected trusted testers	Model card reports simulation and on-robot evaluation, but public task denominators and replication are limited

Access matters because a public reasoning API is different from downloadable VLA weights, a deployable on-robot package, or a supported commercial robot product. The launch page says the whole-body and on-device VLAs remain partner programs. The ER 2 model card lists Google AI Studio and the Gemini API as distribution channels. The On-Device 2 model card limits distribution to trusted testers.

What the whole-body VLA demonstrates

The central technical change is the move from upper-body tabletop manipulation to coordinated movement from feet to fingertips. In Google’s examples, Apollo 2 walks, crouches, stretches, and manipulates objects while following language instructions. One watering-can sequence requires walking to a table, grasping the object, moving to shelving, and placing it in a lower bin (launch source).

Google says the same checkpoint controls three embodiments:

• Apollo 2 with 22-degree-of-freedom SharpaWave hands;
• Apollo 2 with Inspire hands;
• Franka Duo with Robotiq grippers.

That is meaningful cross-embodiment evidence because the comparison does not rely on a separately trained checkpoint for every listed hand and robot configuration. It also expands the action space from arm manipulation into locomotion, balance, body placement, and two-handed object handling.

The shared checkpoint does not make the complete execution stack body-independent. Our guide to the embodiment gap explains why sensor configuration, calibration, controller alignment, physical adaptation, contact, safety, and recovery should still be reported for every target robot.

The launch supports several bounded conclusions:

• one VLA checkpoint can produce whole-body and manipulation actions on the disclosed platforms;
• the model handles hands and parallel grippers with different kinematics;
• Apollo 2 can execute multi-stage first-party demonstrations under Google’s setup;
• fine multi-finger manipulation remains harder than whole-body and gripper categories;
• Google itself says movement speed and precision still need improvement.

It does not disclose enough to convert those examples into a general success rate. The launch describes category bars as averages over multiple tasks, and reports some individual multi-finger task results, but the public article does not provide a complete manifest of attempts per task, resets, human interventions, scene novelty, checkpoint selection, or failure videos.

For a systematic way to read those missing fields, use our robot benchmark evidence card.

ER 2 is the planner, not the motor controller

Gemini Robotics ER 2 is a vision-language model based on Gemini 3.5 Flash. Its model card lists text, image, video, and audio inputs, a context window up to 128,000 tokens, and text output. It does not directly output robot joint commands.

Instead, developers expose lower-level capabilities as tools. Those tools can be a VLA, navigation API, manipulator command, simulator, or even a human teleoperator. ER 2 observes the task, chooses the next tool call, monitors video, and decides when to continue, retry, stop, or ask for help.

This architecture creates a useful separation:

• the reasoning model decides what should happen next;
• the VLA or robot API decides how to produce the motion;
• conventional controllers and hardware safety systems still manage collision avoidance, balance, force, and emergency behavior.

Google’s ER 2 overview reports evaluations across real VLA, simulated VLA, and human teleoperation control modes. It also shows a Boston Dynamics Spot fetching an object through navigation and manipulator APIs. These examples establish orchestration across different execution backends. They do not show that ER 2 makes each backend autonomous or equally reliable.

Progress tracking has numbers, but they need context

Long tasks fail when a system cannot tell whether a step has started, completed, or gone wrong. Google evaluates two temporal capabilities:

• progress classification: assign each video frame to one of five completion bands;
• moment finding: identify the frame where a critical event occurs.

The ER 2 technical overview reports 57.4% accuracy for progress classification. For moment finding, it reports 91.3% accuracy and 0.96 seconds mean absolute distance. Google also claims four times the execution speed of a compared larger model category.

These are model-level video and orchestration metrics. They are not robot task-completion rates. A model can identify the right transition point while a gripper still slips, a navigation system fails, or a safety controller stops the machine. Conversely, a complete task can succeed despite an imperfect five-bin progress label.

A useful deployment report would connect the temporal metrics to end-to-end outcomes: tasks attempted, tasks completed, incorrect transitions, unnecessary retries, unsafe tool calls prevented, human clarifications, wall-clock time, and recovery success.

Multi-robot collaboration is orchestration evidence

Google demonstrates Apollo 2 and Franka Duo dividing work through shared task reasoning. ER 2 provides the semantic coordination layer while each robot retains its own execution interface.

This is stronger than showing two independently scripted robots in the same video because the stated architecture handles task assignment, handoff, and progress. It remains first-party demonstration evidence. Public materials do not disclose a trial denominator, communication failure rate, deadlock behavior, intervention rate, or comparison with a conventional fleet orchestrator.

The defensible interpretation is that ER 2 can coordinate heterogeneous robots in the demonstrated workflows. It is not yet evidence that multi-robot collaboration improves factory throughput or reliability.

On-Device 2 narrows the local-inference claim

On-device inference can reduce network dependence, preserve a tighter latency budget, and allow a robot to operate where connectivity is unreliable. Google says On-Device 2 is based on its on-device Gemma models and accepts language, images, and proprioception before outputting numerical actions.

The launch says a new bi-arm embodiment can be adapted with a few hours of data, typically fewer than 200 examples. The model card says the system was evaluated in simulation and on robots for scene, instruction, and action generalization. It reports faster learning and a higher final success rate than On-Device 1 on platforms introduced during post-training.

The limitations are unusually important:

• public weights or a general download are not listed;
• access is limited to trusted testers;
• the model card describes standing bi-arm manipulation as its primary evaluated scope;
• mobile and whole-body risks are explicitly outside that scope;
• the card says out-of-distribution tasks and high-degree-of-freedom control remain limitations;
• public materials do not provide a complete hardware latency table, power use, thermal behavior, or every adaptation denominator.

The result supports fast adaptation under Google’s tested conditions. It does not establish universal few-shot transfer to any robot.

Safety benchmarks are not robot certification

Google introduces ASIMOV-Agentic, a benchmark for whether a reasoning model can enforce safety constraints, reject unsafe VLA tool calls, recognize infeasible tasks, and request human clarification under uncertainty. The company also reports improvements in safety-instruction following and human-proximity monitoring.

Those are relevant semantic safety tests. They address failures that conventional collision controllers may not understand, such as an unsafe instruction or a task that should be declined. They do not replace system-level engineering.

The ER 2 model card is explicit. Users should not use the robotics models for safety-critical applications, including settings where a malfunction could cause death, injury, or property damage. The On-Device 2 card recommends layers that include embodied reasoning, collision-free motion planning, balance and force control, and hardware-specific functional safety.

A benchmark score therefore does not certify Apollo 2, Spot, Franka Duo, or any other complete system. Certification and safe deployment depend on the robot, task, tooling, workspace, risk assessment, controllers, validated limits, and operating procedures. Our robot safety standards guide explains why AI behavior and functional safety remain different evidence classes.

What is still missing

A production-grade evidence package would publish:

1. a task manifest for every whole-body, gripper, and multi-finger category;
2. attempts, successes, timeouts, resets, exclusions, and confidence intervals per task;
3. motion speed and complete observation-to-action latency distributions;
4. intervention, teleoperation, and safety-stop rules;
5. new-scene, new-object, and new-embodiment split definitions;
6. adaptation examples, duration, and selection rules for On-Device 2;
7. failure categories for planning, perception, locomotion, grasping, and tool orchestration;
8. independent replication on hardware outside Google’s partner setup;
9. customer-side runtime, throughput, uptime, maintenance, and support data;
10. commercial terms and a supported deployment boundary.

Until those fields exist, the Apollo 2 clips belong in the demonstrated category, not the production deployed category. That distinction also applies in our humanoid robot availability tracker.

Evidence verdict

Claim	Classification	Confidence	Why
Gemini Robotics 2 controls whole-body Apollo 2 movement and manipulation	First-party demonstration and benchmark	Medium-high	Google shows multiple platforms and one stated checkpoint, but complete task denominators are not public
ER 2 can orchestrate tools from continuous multimodal input	Public model capability and first-party evaluation	High for access, medium for broad robot reliability	API access and architecture are documented; end-to-end field evidence remains bounded
ER 2 supports heterogeneous robot collaboration	First-party demonstration	Medium	Apollo 2, Franka Duo, and Spot examples are public, but no independent or operational trial log is disclosed
On-Device 2 adapts with fewer than 200 examples	Company-reported evaluation	Medium	The model card supports the intended capability, while complete task and hardware details remain limited
The stack is ready for safety-critical operation	Not supported	High	Google’s own model card warns against safety-critical use
The launch proves a commercial Apollo 2 deployment	Not supported	High	No recurring customer operation, throughput, uptime, or commercial record is included

Bottom line

Gemini Robotics 2 matters because it joins three previously fragmented layers: whole-body action, continuous embodied reasoning, and local cross-embodiment adaptation. Google also documents the layers separately enough to show where a result belongs.

That separation should guide how the launch is read. The VLA demonstrations are not ER 2 benchmark scores. ER 2 API access is not public access to the VLA. On-device adaptation is not evidence of whole-body humanoid safety. Apollo 2 clips are not customer deployment records.

The release advances the model stack and the quality of cross-embodiment demonstrations. The next evidentiary step is not a more polished video. It is task-level trial reporting, independent replication, and sustained customer operation.

Frequently asked questions

Is Gemini Robotics 2 publicly available?

Gemini Robotics ER 2 is available through Google AI Studio and the Gemini API, with a private preview on Gemini Enterprise Agent Platform. The whole-body VLA is available to early-access partners, while On-Device 2 is distributed to selected trusted testers. These are three different access states.

Does Gemini Robotics 2 control a whole humanoid?

Google demonstrates one Gemini Robotics 2 checkpoint controlling Apptronik Apollo 2 configurations across walking, crouching, reaching, and manipulation. This is first-party demonstration and benchmark evidence. Public sources do not provide complete task-level trial counts, intervention rules, or independent replication.

What is the difference between Gemini Robotics 2 and ER 2?

Gemini Robotics 2 is a vision-language-action model that outputs motor actions. ER 2 is a high-level vision-language model that interprets continuous video, plans steps, calls tools, tracks progress, and delegates execution to a VLA or robot API.

Does On-Device 2 work without an internet connection?

Google designed On-Device 2 for local robot inference where network latency or connectivity is undesirable. The public model card says it is intended primarily for standing bi-arm manipulation and that mobile or whole-body risks are outside its current evaluation scope.

Is Gemini Robotics 2 safe for safety-critical work?

No such conclusion is supported. Google's ER 2 model card tells users not to use the robotics models for safety-critical applications. ASIMOV-Agentic and the reported human-proximity evaluations measure semantic safety behavior, not functional-safety certification of a complete robot system.

Does the launch prove a commercial robot deployment?

No. The retained evidence consists of first-party model announcements, demonstrations, benchmarks, model cards, developer access, and partner hardware. It does not establish recurring customer operation, production throughput, uptime, intervention rate, or commercial return.