Google DeepMind drops Gemini Robotics 2.5

While software agents dominated developer headlines, DeepMind's progress on embodied AI demonstrates that foundation models are crossing the reliability threshold for industrial physical deployment. Generalist robotics is transitioning from manual teleoperation toward zero-shot visual reasoning and simulation-accelerated policy learning.

• –Real-world factory validation: Deploying on Apollo 2 humanoids at Mercedes-Benz plants with a 96% kitting success rate proves VLAs can handle high-value, repetitive manufacturing workflows without extensive custom engineering.
• –Hardware-agnostic motor control: By unifying control across parallel grippers like Franka Duo and five-finger dexterous hands like SharpaWave, DeepMind is positioning Gemini as a universal operating system for robotic manipulation.
• –Simulation flywheels with Argon: Training with digital twins generated by Gemini 4 Argon closes the physical robotics data bottleneck, allowing thousands of simulated practice runs before touching hardware.
• –Real-time safety steerability: Inference-time reasoning guidance boosts obstacle avoidance by 33% and human-aware safety by 43%, addressing the primary barrier preventing autonomous humanoids from sharing floor space with human workers.
• –Open safety benchmarks: The release of ASIMOV-Jailbreak sets a much-needed evaluation standard for physical agentic safety, where prompt injection risks tangible physical damage rather than just offensive text.