On Tuesday, Google Deepmind released a new language model called Gemini Robotics On-Device, which allows robots to perform tasks locally without the need for an internet connection.
Based on the company’s previous Gemini Robotics model released in March, the Gemini Robotics On-Device can control the movement of the robot. Developers can control and fine-tune the model to suit a variety of needs using natural language prompts.
In the benchmark, Google claims that the model runs at a level closer to the cloud-based Gemini Robotics model. The company says it is superior to other device models in its general benchmarks, but these models are not named.
In the demo, the company showed that the robot running this local model was doing things like unzipping bags or folding clothes. Google says that while the model was trained by Aloha Robots, it was later adapted by Apptronik to work with Bi-Arm Franka FR3 robots and Apollo Humanoid Robots.
Google claims that the Bi-Arm Franka FR3 has managed to tackle previously “unseen” scenarios and objects, just like it does assemblies on an industrial belt.
Google Deepmind has also released the Gemini Robotics SDK. The company said the developers could display demonstrations of 50-100 tasks on the robot and train them on new tasks using these models of the Mujoco Physics simulator.
Other AI model developers are also dipping their toes into robotics. Nvidia is building a platform for creating basic humanoid models. The embracing face not only develops open models and datasets for robotics, but also works with robots. And Korean startup RLWRLD from Mirae Asset is working on creating a basic model of the robot.
