Google DeepMind has launched Gemini Robotics 2, an advanced AI model for full-body robot control, enhancing capabilities in task planning and execution.
Google DeepMind has unveiled its latest innovation, Gemini Robotics 2, an advanced AI model engineered for full-body control of robots. This model is complemented by the ER 2 solution, which enhances planning and coordination capabilities across multiple robotic devices.
Gemini Robotics 2 is designed to interpret both text and visual cues to execute motor functions, making it adaptable to a range of robotic designs, from small desktop manipulators to full-sized humanoids.
In a significant upgrade from its predecessor, this model now enables robots to perform complex movements such as walking, squatting, bending, stretching, and lifting objects in tight spaces.
During demonstrations, the humanoid robot Apptronik Apollo 2, driven by Gemini Robotics 2, showcased its abilities by picking up a watering can, retrieving a baseball glove from a shelf, and identifying specific items.
DeepMind has also improved the dexterity of robotic hands, now featuring 22 degrees of freedom, allowing robots to perform tasks like tying knots, packaging, and twisting light bulbs.
Additionally, the ER 2 model, described as the "high-level brain" for robots, can plan multi-step tasks, track task completion, and coordinate actions among various robots. In a demonstration, Apollo 2 directed a dual-arm robot to sort tools while cleaning a garage.
Another notable release is Gemini Robotics On-Device 2, a local model that functions without internet access and can adapt to new robotic setups with minimal training time, needing fewer than 200 examples.
Currently, Gemini Robotics 2 is in private preview, while ER 2 is available through the Gemini API and Google AI Studio. Selected partners are testing On-Device 2.
Notably, in July, Humanoid introduced a new method for training robots on real-world production tasks, indicating ongoing advancements in the field.
