Google’s DeepMind division, on Tuesday, June 24, released a new large language model called Gemini Robotics On-Device that runs locally on robotic devices. In a blog post, Google says that the new AI model has been optimised to efficiently run on the robot and shows “strong general-purpose dexterity and task generalisation.”
The new offline AI model builds on the company’s Gemini Robotics model, which the tech giant unveiled earlier this year in March. The Gemini Robotics On-Device model can control a robot’s movement and, like ChatGPT, can understand natural language prompts. Since it works without an active internet connection, Google says it is really useful for latency-sensitive applications or in areas where there is zero connectivity.
Designed for robots with two arms, Google explains that Gemini Robotics On-Device is engineered in such a way that it requires “minimal computational resources” and can complete highly dexterous tasks like folding clothes and unzipping bags, to name a few. You can watch the new AI model in action in the video below.
Compared to other on-device alternatives, Google claims that Gemini Robotics On-Device outperforms the competition when it comes to completing complex multi-step instructions and challenging out-of-distribution tasks. Coming to the benchmarks, it looks like Google’s new offline model comes close to its cloud-based offering. Initially trained for work with ALOHA robots, the company says its new model has been adapted and successfully worked on a bi-arm Franka FR3 robot and an Apollo humanoid as well.
The tech giant said that on the bi-arm Franka FR3, the model was able to follow general-purpose instructions and handle previously unseen objects and scenes, like executing industrial belt assembly. As for Apollo, the model was allowed the humanoid robot manipulate different and unseen objects in a general manner. Developers can try out Gemini Robotics On-Device using the software development kit (SDK).
Google isn’t the only tech giant working on AI models for robots. At GTC 2025, NVIDIA unveiled Groot N1, an AI model for humanoid robots, while Hugging Face is working on developing its very own robot powered by an in-house developed open-sourced model.
© IE Online Media Services Pvt Ltd
0
0
Average Rating