Fine motor tasks in robotics hovered at success rates of 30% to 40%, but the latest deployment on active automotive lines reached a 96% success rate. Google DeepMind introduced Gemini Robotics 2.5, describing it as the company's most capable vision-language-action model to date. The release marks an operational pivot for industrial robotics, as high failure rates previously turned humanoid hardware pilots into capital drains rather than scalable manufacturing assets.
The core issue facing plant operators is straightforward: how much does it actually cost to integrate new vision-language-action software onto automotive assembly lines? According to market assessments of industrial automation integration, deploying bespoke software solutions typically demands hundreds of hours of custom path-planning and sensor calibration per workstation. However, the growth of model accuracy to 96% on an active conveyor shifts robotic arms from experimental balance-sheet line items into operational workflow optimization. If foundational systems can absorb varied sorting and handling tasks without bespoke programming for every physical variation, the economics of deployed machinery change dramatically by amortizing software engineering costs across multiple factory use cases.
Verification on the Automotive Line
To measure dependable performance in production conditions, Google DeepMind partnered with Apptronik and Mercedes-Benz to deploy Apollo 2 humanoid robots on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. In these pilots, the humanoid hardware is equipped with the five-finger, 22 degree-of-freedom SharpaWave hand to manage automotive kitting.
During a real-world proof-of-concept on Mercedes-Benz production lines, our models helped Apptronik’s Apollo 2 humanoid pick and pack automotive parts with a 96% success rate.
As Google DeepMind stated, the model directs the robots to pick up metal automotive brackets from supply bins and place them into sequencing trays while autonomously correcting for gaps. By offloading this monotonous prep work, human technicians can focus on complex assembly tasks. Compared to previous baselines, Gemini Robotics 2.5 delivers a 30% to 100% higher success rate depending on the task type, while finishing operations 20% to 30% faster.
Physical Safety and Inference-Time Steering
Broader deployment across active plants hinges on whether generalized models can overcome customization overhead without driving up capital expenditures. The system addresses floor variability using inference-time reasoning guidance to navigate around physical hazards in real time. Across developer evaluations, this safety feature improved safe task execution around obstacles by 33% and near plant personnel by 43%.
In our evaluations, this capability boosted safe task completion around obstacles by 33% and around nearby people by 43%.
Gemini Robotics 2.5 is rolling out to robotics partners after testing across over 200 distinct tasks that evaluated task breadth, performance, and complexity. Will Mercedes-Benz and Apptronik find that a 96% success rate in Kecskemét and Sindelfingen eliminates enough custom engineering expenses to justify scaling Apollo 2 across remaining assembly lines?
The catch
Plant managers eliminate bespoke coding while picking speeds jump up to 30% alongside floor staff. Will a four percent error rate across lines in Kecskemét and Sindelfingen outweigh the savings from ditching hundreds of hours of workstation-level calibration?
Our stories: Google Gemini Robotics-ER 1.6: AI Agents in Industrial Tech (June 3); Google Gemini Robotics 2: A New OS for Humanoid Robots (August 4).