AI
DeepMind's Gemini Robotics 2.5 Tackles Complex Tasks with Near-Human Dexterity
Google's latest robotic AI platform demonstrates significant improvements in fine motor control and real-world factory performance, though matching human reliability remains elusive.
Ars Technica
Google DeepMind has unveiled the next iteration of its Gemini Robotics platform, bringing machines closer to human-level performance in physical tasks. The 2.5 release marks a substantial leap forward in robotic dexterity, with systems now capable of handling delicate and complex operations that have long challenged artificial intelligence.
What's Changed#
The upgraded system improves upon its predecessor through two core layers of computational reasoning. Gemini Embodied Reasoning—the vision-language component—takes human instructions and processes live video feeds to devise task strategies. Working alongside it, the vision-language-action model generates the precise motor commands that actually execute physical movements. This pairing allows robots to handle operations requiring fine motor control far more effectively than before.
The performance gains are measurable: the system achieves 30 to 100 percent higher success rates on laboratory tasks while running 20 to 30 percent faster. Notably, robots equipped with this upgrade can now tie bows using five-fingered hands—a feat that highlights progress with deformable objects, which have historically stumped robotic systems.
From Lab to Factory Floor#
The real test came beyond controlled environments. Google partnered with Apptronik and Mercedes-Benz to deploy Gemini Robotics 2.5 in automotive manufacturing. The Apollo 2 humanoid robots handled repetitive parts-kitting tasks in European factories, achieving a 96 percent success rate. This field performance suggests the platform can navigate real-world variability and uncertainty more reliably than earlier versions.
However, a 96 percent success rate still falls short of what human workers achieve on identical tasks. According to DeepMind director Kanishka Rao, closing that remaining 4 percent gap presents a "steep road" ahead—a reminder that high-stakes production environments demand near-perfect consistency before robots can operate independently.
Preparing for Workplace Integration#
Alongside the robotics upgrade, DeepMind has released an open-source safety benchmark, signaling the organization's recognition that workplace deployment requires rigorous safeguards as robotic systems become more capable and prevalent in shared human environments.

