Google introduces Gemini Robotics 1.5 for robots

Google DeepMind has introduced Gemini Robotics 1.5, a system designed to help robots do more than receive commands. They can understand a situation, plan several steps and act in the physical world.

The update includes two models that work together. One decides what to do, while the other turns those decisions into specific movements. The aim is to bring robots closer to everyday tasks that remain difficult because they require context, tools and the ability to adjust a plan as they go.

Two models with the same goal

Gemini Robotics-ER 1.5 acts as the coordinator. It is a vision-and-language model that analyzes its surroundings, understands instructions, creates detailed plans and can use digital tools such as Google Search or functions defined by a developer.

Gemini Robotics 1.5 handles plan execution. It is a vision-language-action model, which means it turns what it sees and the instructions it receives into commands for the robot's motors.

For example, when asked to sort waste according to local rules, the system could first search for the area's recycling regulations, identify the objects in front of it, decide which bin each one belongs in and complete the task physically.

The robot plans before it moves

Models of this kind typically translate an instruction into movements. Gemini Robotics 1.5 adds an analysis phase before acting: it can break a long task into simpler steps and consider both the overall goal and the movements needed to achieve it.

In a task such as sorting clothes by color, it first interprets what the instruction means. It then decides which item to pick up, where to place it and how to approach it for a better grip. This process is intended to make the robot more resilient to changes in its surroundings and better able to handle tasks it did not see exactly during training.

Google also says the model can explain its decisions in natural language. That could make human supervision easier, although an explanation does not replace the safety measures that control the robot's movements.

Learning with different types of robots

Robots have different bodies, sensors and movement capabilities. As a result, a skill learned on one model often needs adjustments to work on another.

According to Google, Gemini Robotics 1.5 can transfer movements between different platforms without starting from scratch for each one. The company says tasks trained with the ALOHA 2 robot also work on Apptronik's Apollo humanoid and the two-armed Franka robot, and vice versa.

If this capability holds up beyond controlled demonstrations, it could reduce the time needed to teach new skills to each robot. That is one of the obstacles separating prototypes that can perform one task from robots that are truly versatile.

Availability and safety

Gemini Robotics-ER 1.5 is already available to developers through the Gemini API in Google AI Studio. Gemini Robotics 1.5, by contrast, is currently available to a selected group of partners.

Google says Gemini Robotics-ER 1.5 achieves leading results on 15 academic benchmarks and in internal evaluations related to spatial understanding. It has also updated ASIMOV, its test suite for evaluating the semantic safety of robots, with new data, videos and question types.

The safety approach combines reasoning about the task before acting with human-interaction policies and low-level systems such as collision prevention. In practice, that means the robot must assess whether an action is appropriate while also relying on specific controls to avoid hitting people or objects.

For you, the important change is not that robots will handle any household task tomorrow. It is that Google is trying to move from robots that respond to isolated commands to systems capable of finding information, organizing a mission and carrying it out step by step. The next thing to watch is how these models behave outside test environments, where objects, instructions and risks are rarely perfectly controlled.