Gemini Robotics 2.5 hits 96% success in Mercedes-Benz pilot

Apptronik’s Apollo 2 humanoids achieved a 96% success rate sorting car parts on active Mercedes-Benz production lines in Hungary and Germany (Scott D. Clary from 4IR). Google DeepMind split this work between two models: Gemini Robotics ER, which managed the workflow sequence, and Gemini Robotics 2.5, which controlled the robot's physical movements (TLDR Hardware). The system functions by translating live video and text commands into precise robot motions, allowing the humanoids to pull metal brackets from bulk bins and set them into specific trays for vehicle assembly (Sifu Yik's Substack).

Gemini Robotics 2.5 is now rolling out to Google’s robotics partners, including Agile Robots and Boston Dynamics, to support diverse physical tasks ranging from serving food with tongs to laboratory pipetting. Google DeepMind reports that this model improves task success rates by 30% to 100% over earlier baselines and increases task execution speed by 20% to 30%. Safety features boosted by Gemini 4 Argon also increased obstacle and human interaction safety by 33% and 43% respectively (TLDR Hardware).

The Mercedes-Benz work remains a proof of concept as the broader industry experiments with whole-body control to further improve agility. Boston Dynamics recently added wheels to its Atlas humanoid to test transitions between rolling and legged motion, contributing to the ongoing debate regarding the most effective movement architecture for unstructured environments.