Introduction
Google DeepMind’s current Gemini Robotics family separates robot action models from embodied-reasoning models. Gemini Robotics 2 is the cloud VLA for whole-body and dexterous control. Gemini Robotics-ER 2 handles higher-level spatial and temporal reasoning, tool orchestration and multi-step planning. Gemini Robotics On-Device 2 is the local VLA option for supported robot integrations. The original 2025 models remain useful historical references, but current capability and access claims should use the named 2026 version. Public benchmarks come from different robots and tasks, so the results stay attached to the model, embodiment and evaluation setup that produced them.
Key findings
- The current family separates Gemini Robotics 2 for robot actions, ER 2 for high-level embodied reasoning and On-Device 2 for local execution.
- Real-robot evidence covers controlled manipulation and multiple embodiments.
- Access differs by model: ER 2 is available through Google’s developer interfaces, while the VLA models use partner or early-access programs.
- Failures include incorrect grounding, depth error, action drift, unseen materials and weak recovery after a failed grasp.
- Practical uses include manipulation policies, spatial grounding, grasp proposals and planning support.
Gemini Robotics From Embodied Reasoning to Robot Actions - evidence comparison
The table uses source-backed fields and leaves non-comparable or undisclosed information visible.
| System, category or question | Verified evidence | Interpretation or limitation |
|---|---|---|
| Gemini Robotics | VLA action output | Manipulation on supported robots Direct control model |
| Gemini Robotics-ER | Spatial reasoning | Points, grasps and 3D outputs Needs controller |
| On-Device | Local VLA execution | Low-latency adaptation Access limited at launch |
| Partner integrations | Robot-specific stacks | Apollo, Franka and others Evidence differs |
Evidence classification
- Confirmed by official technical documentation specifications, architecture or access stated by the responsible organization.
- Confirmed by a research paper result reported under a defined experiment, without implying deployment.
- Demonstrated on a real system a physical robot or product performed the documented sequence.
- Company claim without independent verification numerical or operational statement supplied by the company.
- Public evidence insufficient version, control mode, duration, trial count or operating conditions are missing.
Definition and scope
The original 2025 Gemini Robotics release was built on Gemini 2.0 for physical tasks. The original release separated a vision-language-action model producing robot actions from Gemini Robotics-ER, which produces spatial outputs for downstream controllers. Gemini Robotics is software, not a physical robot, and it does not make every partner machine autonomous. Each integration still needs robot control, calibration, safety and task evaluation.
How the complete pipeline works
Images, language and robot state enter multimodal encoders. The VLA grounds instructions, predicts action chunks and executes in closed loop. ER can produce points, grasps, trajectories, correspondences and 3D boxes.
For Gemini Robotics, closed-loop execution means observing the result of each command before the next decision. The high-level component described here does not replace robot-specific motor control, collision handling or independent safety limits.
Key systems, products and technical evidence
The report documents ALOHA 2, bi-arm Franka adaptation and specialization to Apptronik Apollo. On-Device was designed for local inference and reported adaptation from 50 to 100 demonstrations. The systems are not treated as interchangeable.
Evidence from real systems
Real-robot evidence covers controlled manipulation and multiple embodiments. Access has used trusted testers and developer programs; official pages reviewed did not establish open weights. Real-system evidence is separated from simulation, internal testing, controlled public demonstrations, pilots and commercial deployment.
For Gemini Robotics, the strongest report would name the exact version, task boundary, environment, control method, duration, trial count, intervention rate and recovery behavior. The current public record for Gemini Robotics, Gemini Robotics-ER does not provide every field, so the article limits each conclusion to the documented setup.
Comparison method and engineering tradeoffs
The Gemini Robotics comparison uses only fields that can be traced to the cited records. Missing values stay visible instead of receiving estimated scores.
Failure modes and misleading interpretations
Reporting can create a second failure layer around Gemini Robotics. The fact-check therefore labels documentation, real-system evidence, controlled demonstrations, company claims and insufficient evidence separately.
Practical applications and current maturity
Practical uses include manipulation policies, spatial grounding, grasp proposals and planning support. Deployment requires runtime monitoring. These uses are credible only within the documented task, robot and environment.
A team adopting Gemini Robotics should request the exact interfaces and evidence its application needs.
Open problems and recommendations
Open questions include current API scope, pricing, training-data disclosure, long-horizon recovery and support for additional robot embodiments. Public evaluations still use different robots and tasks, so common protocols and unedited trial distributions are needed for direct comparison.
Future Gemini Robotics releases should publish versioned sensor layouts, action spaces, control rates, training or adaptation steps and complete evaluation distributions.
Limitations and missing information
- Benchmarks from different robots, versions, environments or control modes are not directly comparable.
- Company-reported metrics are not independently audited unless a separate primary record establishes the same result.
- Code, weights, prices, model versions, APIs and commercial availability can change after publication.
- Long-duration reliability, intervention frequency and complete failure distributions are rarely published.
Conclusion
The current Gemini Robotics family separates action generation, embodied reasoning and on-device execution into named models with different access paths. Public real-robot evidence covers controlled manipulation on multiple embodiments, while open weights are not documented for the Gemini action models. Direct comparison still requires the exact model version, robot, task, trial count, intervention rate and recovery behavior.
Frequently asked questions
What is Gemini Robotics?
The original 2025 Gemini Robotics release was built on Gemini 2.0 for physical tasks. The original release separated a vision-language-action model producing robot actions from Gemini Robotics-ER, which produces spatial outputs for downstream controllers.
How does Gemini Robotics work?
Images, language and robot state enter multimodal encoders. The VLA grounds instructions, predicts action chunks and executes in closed loop. ER can produce points, grasps, trajectories, correspondences and 3D boxes. In practice, calibration, latency, action scaling and feedback determine whether the pipeline remains stable.
What is the strongest real-world evidence?
Google publishes real-robot manipulation evidence for Gemini Robotics action models and spatial-reasoning evaluations for Gemini Robotics-ER. These results come from different tasks and robot setups, so they should remain attached to the published evaluation conditions.
Sources and methodology
Sources for Gemini Robotics were checked on July 11, 2026. The review prioritized the official records from Google DeepMind, Open X-Embodiment Collaboration, NIST, plus primary papers, repositories, model cards, product pages or filings where applicable.
For Gemini Robotics, evidence is sorted by test setting and control mode simulation is kept apart from physical trials, teleoperation from autonomous execution, and announced access from a system that can actually be obtained or deployed.
The review keeps action models, embodied-reasoning models and on-device variants separate because their inputs, outputs, access paths and evaluation conditions differ.
Related TechniaHQRobot guides
Source review notes
Review date recorded July 11, 2026
Supported by the cited sources
- Real-robot evidence covers controlled manipulation and multiple embodiments.
Not confirmed or incomplete
- Failures include incorrect grounding, depth error, action drift, unseen materials and weak recovery after a failed grasp.
- Company-reported metrics are not independently audited unless a separate primary record establishes the same result.
- Long-duration reliability, intervention frequency and complete failure distributions are rarely published.
Likely to change quickly
- Prices, model versions, APIs, software access and commercial availability.
- Production, customer pilots, deployments and repository maintenance status.
Share this article
Share the current TechniaHQRobot article page.
Continue reading
Open the latest robotics reporting, Physical AI analysis and hardware notes.