Google DeepMind has launched Gemini Robotics On-Device, an advanced vision-language-action (VLA) model designed for offline use on robots. This model is smaller, faster, and optimized for bi-arm robots, enabling real-time tasks like folding clothes and industrial automation without relying on cloud support. It has been tested on various platforms, including the Franka FR3 and the humanoid Apollo robot, demonstrating excellent adaptability and response to natural language commands.
To facilitate development, the Gemini Robotics SDK allows developers to experiment with physics simulations and fine-tune the model using just 50-100 demonstrations. Benchmark results highlight the model’s strong performance in generalization, instruction-following, and multi-step actions, making it suitable for various applications. Safety features are integrated, including content safety through a Live API and oversight from dedicated risk assessment teams. This launch represents a significant step towards enhancing robotics AI’s adaptability and application potential, with broader availability expected later.
Source link
