Google DeepMind Unveils Gemini Robotics 2.5 for Enhanced Physical Dexterity

Google DeepMind has introduced Gemini Robotics 2.5, its latest vision-language-action model aimed at expanding the range of skills robots can acquire and accelerating their learning speed in unstructured physical environments. According to DeepMind, the model works across hardware platforms, including Apptronik's Apollo 2 humanoid and the Franka Duo system. Evaluations across 200 distinct manipulation tasks showed a 30% to 100% higher success rate and a 20% to 30% faster execution time relative to prior baselines. In an industrial proof-of-concept conducted with Mercedes-Benz and Apptronik in Germany and Hungary, an Apollo 2 robot achieved a 96% success rate performing automotive parts kitting on active production lines. Additionally, Gemini Robotics 2.5 introduces inference-time reasoning guidance to steer around obstacles and people, alongside an open-source evaluation benchmark titled ASIMOV-Jailbreak to test robotic resistance against unsafe requests.

Key takeaways

• Google DeepMind introduced Gemini Robotics 2.5, expanding manipulation skill breadth, dexterity, and learning speed across unstructured real-world tasks.
• During automotive kitting trials on Mercedes-Benz assembly lines, Apptronik's Apollo 2 humanoid achieved a 96% success rate using the model.
• Evaluations across 200 tasks showed 30% to 100% higher task success rates and 20% to 30% faster execution compared to previous baselines.
• DeepMind added real-time hazard steering guidance and released the ASIMOV-Jailbreak benchmark to test robustness against unsafe physical instructions.