September 25, 2025, as humanoid robots have become a market focus, Google DeepMind has unveiled two robot AI models: Gemini Robotics‑ER 1.5 and Gemini Robotics 1.5, aimed at giving robots stronger perception, reasoning, planning and execution capabilities to advance them from passive execution toward active thinking plus action. Alongside the release, DeepMind has opened a developer preview of the Gemini Robotics‑ER model.
1. Model design and roles: separation of thinking and acting
DeepMind designed the two models to play complementary roles:
- Gemini Robotics‑ER 1.5 (Embodied Reasoning): a high‑level reasoning model specialized in spatial reasoning, planning, task decomposition and tool use in the physical world. It can interpret visual input, environmental state and natural language instructions, converting these into multi‑step action plans.
- Gemini Robotics 1.5 (Vision‑Language‑Action, VLA): focuses on translating high‑level plans into concrete actions. Given a plan and visual input, it generates robot control commands to execute physical motions, and can present its chain of thought to improve transparency.
This architecture enables robots to think before acting instead of simply executing single‑step commands. DeepMind states in its blog that the two models working together allow robots to complete multi‑step tasks in more complex environments.
2. Key capabilities and highlights
1. Improved spatial and task understanding
Gemini Robotics‑ER 1.5 shows strong performance on spatial reasoning, goal localization and trajectory prediction, leading on several embodied reasoning benchmarks including ERQA, Point‑Bench and RoboSpatial‑VQA. It has been tuned for high‑level planning, environment understanding and task decomposition.
2. Tool use and online connectivity
A key advance is native support in Gemini Robotics‑ER 1.5 for tool use, including calling Google Search or user‑defined functions. That means if a robot faces an unfamiliar task it can first look up information online and then formulate a plan based on what it finds. For example, a robot could query local waste‑sorting rules or weather conditions to act more compliantly and intelligently on site.
3. Adjustable thinking budget and safety filtering
DeepMind provides a thinking budget control so developers can trade off latency against accuracy: simple tasks can use less thinking time while complex tasks can be allowed more planning time to improve results.
At the same time, the models improve safety filtering, rejecting unreasonable or dangerous operations (such as overload or collision risks) already at the planning stage.
3. Cross‑hardware transfer and generality
A notable capability is cross‑embodiment skill transfer. DeepMind demonstrated transferring skills learned on one robot platform (for example, the ALOHA2 arm) directly to structurally different robots (such as dual‑arm Franka or Apptronik's humanoid Apollo) without full retraining. This enables shared training knowledge across robots and speeds skill expansion.
DeepMind also says these models build on the Gemini series foundation, retaining multimodal understanding (images, text, video, etc.) and generalizing these abilities to physical‑world operation.
4. Availability and deployment strategy
DeepMind states that Gemini Robotics‑ER 1.5 is available to developers in preview via the Gemini API on Google AI Studio; Gemini Robotics 1.5 is currently available only to selected partners. Developers can call the ER model via API for planning and pass the plan steps to the VLA model for execution.
The official blog also provides SDK examples, limited test programs and technical documentation to help teams start building robot applications.
5. Challenges and risks
Despite the significance of the release, both models face several challenges:
- Action precision and execution robustness: moving from planning to physical action still requires overcoming grasping errors, control accuracy and collision details.
- Safety and trustworthiness: real‑world execution errors can cause physical harm; DeepMind introduces safety filters and ASIMOV benchmark evaluations to mitigate these risks.
- Generality versus specialization: although cross‑platform transfer is supported, robots with substantially different configurations may still require targeted fine‑tuning.
- Cost and resource consumption: high‑performance perception plus reasoning models demand significant compute, posing barriers for edge or low‑resource deployments.
- Real‑world validation: current results are largely demos and benchmark tests; deploying in complex, variable settings (household chores, medical assistance, warehousing, etc.) will require time and accumulation of field experience.
Conclusion
DeepMind's release of Gemini Robotics‑ER 1.5 and Gemini Robotics 1.5 represents a substantive step in robotics by integrating thinking into robotic systems. Through a layered architecture, tool use and cross‑embodiment transfer, DeepMind aims to make robots more flexible in perceiving, reasoning, planning and executing complex tasks.
If these models withstand real‑world tests, they could extend general AI from virtual to physical domains. In the coming months, as the developer API spreads, community reproductions appear and application‑level demos emerge, the practical value of this ecosystem will become clearer.