Model positioning
Gemini Robotics-ER is a vision-language model introduced by Google based on Gemini 2.0; ER stands for “embodied reasoning.” It focuses on the spatial understanding needed by robots and can work with a robot’s existing low-level controllers. It differs from Gemini Robotics, which outputs physical actions to directly control robots. (deepmind.google)
Capabilities and uses
- Supports spatial understanding tasks such as object detection, pointing to parts of objects, multi-view correspondence, and 3D object detection.
- Can be used for perception, state estimation, spatial understanding, planning, and code generation; for example, inferring how to grasp a cup by its handle and an approach path from an image of the cup.
- Can perform in-context learning from a small number of human demonstrations. Google states that, in its end-to-end robotic control evaluation, the model achieved a success rate 2–3 times that of Gemini 2.0; this result should not be regarded as general performance across all tasks. (deepmind.google)
Limitations and conditions of use
- At release, Google worked with a small number of trusted testers, including Agile Robots, Agility Robots, Boston Dynamics, and Enchanted Tools; the current specific access conditions for the first-generation model have not been verified. (deepmind.google)
- The model’s generated reasoning, code, or action recommendations may be incorrect. Physical robots still require low-level safety controls tailored to the specific hardware, including collision avoidance, contact force limits, and dynamic stability safeguards. (ai.google.dev)
- Google subsequently released versions including Gemini Robotics-ER 1.5; their features, interfaces, and prices cannot be directly treated as attributes of the first-generation model in this record. (deepmind.google)
Pricing
As of 2026-11-19, the official input or output billing prices, currency, and billing units for the first-generation Gemini Robotics-ER had not been verified. Prices for later versions on the official pricing page should not be applied to the first-generation model. Source consulted: Official Google Gemini API pricing. (aistudio.google.com)
Release date and sources
- Release date: 2025-03-12.
- Official announcement: Gemini Robotics brings AI into the physical world. (deepmind.google)