Google DeepMind's Gemini Robotics On-Device Empowers Robots Offline
Google DeepMind has unveiled Gemini Robotics On-Device, a pioneering language model that empowers robots to perform tasks offline and locally. The model has shown impressive results, matching its cloud-based counterpart and even surpassing other on-device models in general benchmarks.
Initially trained for ALOHA robots, Gemini Robotics On-Device has demonstrated remarkable versatility. It successfully controls the bi-arm Franka FR3, managing assembly tasks it hadn't encountered during training. This adaptability extends to other robots like Apollo and those used by partners such as Apptronik, with over 60 testers involved in various robotic platforms.
Demonstrations showcased robots performing complex tasks like unzipping bags and folding clothes using natural language prompts. Gemini Robotics On-Device's performance in benchmarks is comparable to its cloud-based counterpart, indicating its potential for real-world applications. Google DeepMind is now offering a Gemini Robotics SDK, allowing users to train robots on new tasks using the MuJoCo physics simulator.
Gemini Robotics On-Device's success signals a significant step towards more capable and adaptable robots. Its offline and local task execution, combined with its impressive benchmark results, makes it a promising tool for various industries. With the release of the Gemini Robotics SDK, users can now explore and expand the model's capabilities.
