Gemini Robotics ER 2 Preview: Spatial Reasoning and Tool Orchestration
Google released Gemini Robotics ER 2 Preview on July 30, 2026. This article details its spatial reasoning capabilities, context window, and evaluation steps…
Official Release Details and Model Specifications
Google officially released the Gemini Robotics ER 2 Preview on July 30, 2026. This announcement marks the availability of two new embodied reasoning model endpoints designed specifically for robotics applications. The primary endpoint identified in the official changelog is gemini-robotics-er-2-preview. The release date is distinct from the current publication date of this article, which reflects the information as verified in August 2026.
- Provider: Google
- Model Name: Gemini Robotics-ER 2 Preview
- Input Modalities: Text, Image
- Output Modality: Text
- Context Window: 131,072 tokens
- Max Output Tokens: 65,536
- Open Weights: Not confirmed
Documented Capabilities for Embodied Reasoning
The source facts list several specific functional attributes associated with the gemini-robotics-er-2-preview endpoint. These include advanced spatial reasoning, which allows the model to process spatial data within its input modalities. The model also supports agentic code execution, indicating it can generate or run code as part of its workflow. Additionally, the documentation cites multi-step tool orchestration, suggesting the ability to manage sequences of tool calls. Other listed features include video moment finding, progress classification, and multi-robot coordination. These capabilities are documented as part of the preview release but do not confirm specific performance metrics or guaranteed operational benefits.
Practical Evaluation Procedure for Engineers
To evaluate the Gemini Robotics ER 2 Preview, engineers should design tests that directly target the documented features. First, test spatial reasoning by providing image inputs containing complex geometric arrangements and measure the accuracy of the textual descriptions generated. Second, validate agentic code execution by requesting simple robotic control scripts and compare the output syntax against expected standards. Third, assess multi-step tool orchestration by defining a task requiring multiple sequential API calls and observe if the model correctly orders them. Finally, verify multi-robot coordination by simulating a scenario with two distinct robot agents and check if the model assigns appropriate actions to each. Measure latency and token usage during these tests to establish baseline performance. Compare results against previous internal benchmarks if available, but note that no comparative claims are made in the source material.
Primary sources
Frequently asked questions
What is the context window size for Gemini Robotics ER 2 Preview?
The model has a context window of 131,072 tokens and a maximum output limit of 65,536 tokens.
Are open weights available for this model?
The source facts state that open weights are not confirmed. This does not imply local installation is impossible, only that open weight availability is not established in the provided data.