The industry has found a new hobby: robots. Amazon, Meta, Apple, Tesla, and others are investing heavily in this emerging sector, with the ambition of supplying each household with one (or more!) robots to make our daily lives easier.
What if robots finally understood their environment?
But before that, these robots will need a brain so that they can adapt to their environment and handle unexpected situations. Google has understood this well: its DeepMing division has unveiled Gemini Robotics, an AI model based on Gemini 2.0. It was designed to give robots "intelligence" capable of adapting and interacting in the physical world.
Two versions were presented: Gemini Robotics, simply a multimodal model (VMA for vision-language-action) that gives the robot the ability to understand instructions and perform all kinds of tasks. This version was designed to adapt to new situations, handle unknown objects, and perform tasks never encountered during its training, which Google describes as "generality."
Gemini Robotics must also demonstrate "interactivity," by understanding and responding to instructions in natural language. The model must constantly monitor the environment and adjust its actions accordingly and in real time. Finally, there is also talk of "dexterity": armed Based on this model, robots are able to perform very precise manipulations, such as folding a sheet of paper or putting objects in a plastic bag.
All this allows the robot, for example, to immediately replan its action if an object slips from its hands or is moved. Google also presented Gemini Robotics-ER: a version specialized in spatial reasoning and interaction with the environment. This model improves 3D perception, planning, and code generation to execute precise actions.
When shown a cup of coffee, the robot with the Robotics-ER model can intuitively determine that it should use a two-finger grip to grasp the handle and calculate a safe path to approach it.
One of Google's early partners in this space is Apptronik, which is developing Apollo, a humanoid robot. It is not specified While the latter may integrate Gemini Robotics, the goal of this collaboration is to create robots capable of performing tasks in the real world with this technology. Gemini Robotics can thus be used on several types of robots, including two-limbed robotic arms.
Source: Google
0 Comments