Gemini Robotics
ER 2
Model Card​
​
Model Cards are intended to provide essential information on Gemini models, including known
limitations, mitigation approaches, and safety performance. Model cards may be updated from time to
time; for example, to include updated evaluations as the model is improved or revised. See the Google
DeepMind site for a comprehensive list of model cards.

Published: July 2026

Model Information

Description
Gemini Robotics ER (Embodied Reasoning) 2 is a Vision-Language-Model that enhances Gemini’s
spatial, temporal, and physical reasoning capabilities.

Model dependencies
Gemini Robotics ER (Embodied Reasoning) 2 is based on Gemini 3.5 Flash.

Inputs
Text / Image / Video / Audio (interleaving possible). Context window of up to 128k.

Outputs
Text, with up to 64K token output.

Architecture
Gemini Robotics ER 2 is based on the Gemini 3.5 Flash. For more information about the model
architecture, see the Gemini 3.5 Flash model card.

​
Model Data

Training Dataset
Gemini Robotics ER 2 was trained on Gemini 3.5 training datasets and additional datasets representing
various embodied reasoning tasks. For more information about the training data, see the Gemini 3.5
Flash model card.

Training Data Processing
For more information about the training data processing, see the Gemini 3.5 Flash model card.

​

Implementation and Sustainability

Hardware
Gemini Robotics ER 2 was trained using Google’s Tensor Processing Units (TPUs). See details in the
Gemini 3.5 Flash model card.

The efficiencies gained through the use of TPUs are aligned with Google's commitment to operate
sustainably.

Software
Software JAX (Bradbury et al., 2018), ML Pathways (Dean, 2021).

Distribution
Gemini Robotics ER 2 is distributed in the following channels; respective links shared in line:

    ●​ Google AI Studio
    ●​ Gemini API
    ●​ Gemini Enterprise Agent Platform (private preview)
Our models are available to downstream providers via an application program interface (API) and
subject to relevant terms of use. There is no required hardware or software to use the model. For
AI Studio and Gemini API, see the Gemini API Additional Terms of Service; for Gemini Enterprise
Agent Platform, see Google Cloud Platform Terms of Service. For more information, see Gemini
Model API instructions and Gemini API quickstart.

Evaluation

Approach
See Figure 1-4 in the release post for Gemini Robotics ER 2 evaluation procedures and Figure 5 for
Gemini Robotics Safety evaluation procedures.

Results
See Figure 1-4 in the release post for Gemini Robotics ER 2 evaluation results and Figure 5 for
Gemini Robotics Safety evaluation results.

Intended Usage and Limitations

Benefit and Intended Usage
Gemini Robotics ER 2 represents an advancement to our reasoning-first Gemini Robotics ER
models that enables robots to understand their environments with precision. This model
specializes in reasoning critical for robotics, including visual, spatial, and temporal understanding,
tool orchestration and success detection for physical agents.

Known Limitations
For more information about the known limitations for Gemini Robotics ER 2, see the Gemini 3.5
Flash model card.

Acceptable Usage
For more information about the acceptable usage for Gemini Robotics ER 2, see the Gemini 3.5
Flash model card. Additionally, users are required to use discretion before using the Robotics
Models in a production, commercial, or public environments, and to not use the Robotics Models
for safety-critical applications or work, such as in the following settings: (i) healthcare, (ii)
transportation, or (iii) other areas where safety protocols are vital, and a malfunction could
reasonably foreseeably lead to death, personal injury, or property damage.

Ethics and Safety

Ethical Considerations & Risks
Previous impact assessment and risk analysis work as discussed in (Gemini Robotics 1.5) and references
therein remain relevant to Gemini Robotics ER 2. See Section 6 for information on responsible
development and safety mitigations.

Evaluation
Gemini Robotics ER 2 performs similarly to Gemini 3.5 Flash in terms of compliance with Gemini safety
policies which prevent our AI models from generating specific types of harmful content. It achieves
significant gains on Safety Instruction Following, Human Safety monitoring and other semantic safety
benchmarks, see the safety section in the release post. For new agentic safety and uncertainty
resolution evaluations, please see our safety technical report.