DeepMind updates Gemini Robotics 2.5 to improve factory humanoid performance
The new AI platform helps humanoid robots perform complex manual tasks like sorting automotive parts with a 96% success rate.
The short version
- Gemini Robotics 2.5 improved task success rates by 30 to 100 percent in laboratory testing.
- The system is faster at completing tasks than the previous version.
- Apollo 2 humanoid robots achieved a 96 percent success rate sorting parts in European factories.
- The update includes inference-time reasoning to help robots navigate around physical hazards and people.
Why it matters
This update moves AI robotics from controlled labs to unpredictable industrial environments. It suggests that humanoid robots are becoming viable for complex assembly work that requires human-like dexterity.
A leap in robotic dexterity
Google DeepMind recently introduced Gemini Robotics 2.5, an update to its AI-powered platform for controlling robots. The system combines Gemini Embodied Reasoning with Gemini Robotics to turn live video and human instructions into physical actions. This allows robots to perform tasks that require high levels of manual skill and precision.
During laboratory testing, the new version showed significant improvements over its predecessor. DeepMind reported that the update resulted in a 30 to 100 percent higher success rate for various activities. It also made the robots faster at completing those same tasks.
The company evaluated the system across more than 200 different activities to test its versatility. These tasks ranged from using a pipette to serving food with tongs. One specific example shown by DeepMind involved the Apollo 2 humanoid robot using its five-fingered hands to tie a bow.
Real-world factory testing
The technology is moving beyond the lab and onto active production lines. DeepMind partnered with Apptronik and Mercedes-Benz to test the system in industrial settings. The goal was to see if the robots could handle the specific demands of vehicle manufacturing.
In a proof-of-concept, Gemini-powered Apollo 2 robots were used to kit automotive parts in several European factories. These robots were tasked with sorting and placing parts in the exact order required for each vehicle. The robots performed these actions at plants located in Kecskemét, Hungary, and Sindelfingen, Germany.
The results from these live production lines were highly successful. As reported by NewsCord, the robots achieved a 96 percent success rate during the pilot. Apptronik noted that the pairing of the Apollo 2 hardware with the Gemini Robotics software resulted in very little drop-off in performance.
How the AI manages tasks
DeepMind utilized a two-model setup to manage the robots during the factory pilot. One model, Gemini Robotics ER, was responsible for tracking the overall workflow of the production line. The other model, Gemini Robotics 2.5, was used to control the specific movements of the robot's arms and hands.
The system relies on inference-time reasoning guidance to help the robots navigate their surroundings. This technology allows the AI to steer the robot away from physical hazards in real time. It is designed to help the machine make decisions as it encounters unexpected obstacles.
This reasoning capability significantly improves safety in environments where humans are present. DeepMind reported that safe task completion improved by 33 percent when the robot was near obstacles. Furthermore, the success rate improved by 43 percent when people were in the way of the robot's path.
Safety and benchmarking
Alongside the hardware updates, DeepMind released a new safety benchmark called ASIMOV-Jailbreak. This tool is designed to test whether robotic AI can resist requests that are disguised as unsafe instructions. The benchmark includes 1,500 different prompts and is available on the Hugging Face platform.
The company also introduced an open database of these 1,500 prompts to help the research community test robotic AI safety. These tests are crucial for ensuring that robots do not perform dangerous actions when interacting with humans. The benchmark helps developers identify vulnerabilities in how robots interpret commands.
DeepMind noted that the safety level of their systems reached 98.8 percent after adding Argon, a specific component. This compares to an 82 percent figure for a previous version of Gemini Robotics ER. These benchmarks provide a standardized way to measure how well a robot follows safety protocols.
The road ahead for robotics
While the success rates in the Mercedes-Benz plants are high, the transition to full automation is still ongoing. DeepMind director Kanishka Rao acknowledged that there is still significant work to do. He described closing the remaining gap between current capabilities and perfect performance as a "steep road."
The partnership between DeepMind, Apptronik, and Mercedes-Benz serves as a blueprint for future industrial applications. By proving that robots can perform useful work alongside people, the companies are moving toward a more integrated workplace. The focus remains on making robots capable of handling complex, high-dexterity tasks that were previously difficult for machines.
Future developments will likely focus on refining these success rates in even more diverse environments. As the technology matures, the goal is to reduce the remaining gap in performance. This will determine how quickly humanoid robots can be integrated into standard manufacturing and service industries.
What to watch
- How the 96% success rate translates to full-scale factory production.
- Whether other manufacturers besides Mercedes-Benz adopt Gemini Robotics 2.5.
- How the 'steep road' of remaining gaps will be addressed in the next software update.
Try it
Look up the ASIMOV-Jailbreak benchmark on Hugging Face to see how researchers are testing robot safety.
Sources and links
- ReportingNewsCord: the article we sourced
First reported by NewsCord. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.