Google’s Gemini Robotics 2.5 scores 96% in factory trial after tackling 200 tasks
Google’s Gemini Robotics 2.5 scores 96% in factory trial after tackling 200 tasks
Google DeepMind introduced Gemini Robotics 2.5, a new vision-language-action model designed to accelerate robot learning of physical tasks and improve reliability in real-world settings. In a proof-of-concept trial at Mercedes-Benz production facilities in Hungary and Germany, the model achieved a 96% success rate on Apptronik's Apollo 2 humanoid robots as they autonomously sorted automotive parts into trays for assembly. The system was evaluated across more than 200 distinct tasks, including handling delicate objects and using tools, showing significant improvements over previous baselines. Google also highlighted safety enhancements that increased safe task completion around obstacles and people, alongside the release of an open-source benchmark to test resistance to unsafe instructions. The model is currently being rolled out to Google's robotics partners.
Source: Robotics — Web news discovery · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Control
- Reported autonomousOpen source S2
Reported numbers
Reported success rate
96%
Reported trials
View original evidence
Google DeepMind has introduced Gemini Robotics 2.5, a new model designed to help robots learn physical tasks faster, handle delicate objects, and operate more reliably in real-world environments. The model achieved a 96 percent success rate in a factory proof-of-concept involving Apptronik’s Apollo 2 humanoid robots on Mercedes-Benz production lines. The robots autonomously sorted automotive parts from bulk bins into
Open source S2task count
200
View original evidence
which translates visual information and instructions into physical robot movements. It is currently rolling out to Google’s robotics partners. Model tackles 200 tasks Google evaluated Gemini Robotics 2.5 across more than 200 distinct tasks, including serving food with tongs, tying bows on gift boxes, and transferring liquids with pipettes. These activities test a range of capabilities, from handling delicate objects
Open source S4Reported success rate
30%
View original evidence
to manipulating flexible materials and using everyday tools. Compared with previous baselines, the model achieved success rates that were 30 percent to 100 percent higher, depending on the task. It also completed activities 20 percent to 30 percent faster, according to Google. The model works across different robotic platforms, including Apptronik’s Apollo 2 humanoid and the Franka Duo system, which uses two-fingered
Open source S5Reported success rate
100%
View original evidence
to manipulating flexible materials and using everyday tools. Compared with previous baselines, the model achieved success rates that were 30 percent to 100 percent higher, depending on the task. It also completed activities 20 percent to 30 percent faster, according to Google. The model works across different robotic platforms, including Apptronik’s Apollo 2 humanoid and the Franka Duo system, which uses two-fingered
Open source S5
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
What remains unknown
Not established in the collected evidence: Environment, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
Google DeepMind has introduced Gemini Robotics 2.5, a new model designed to help robots learn physical tasks faster, handle delicate objects, and operate more reliably in real-world environments. The model achieved a 96 percent success rate in a factory proof-of-concept involving Apptronik’s Apollo 2 humanoid robots on Mercedes-Benz production lines. The robots autonomously sorted automotive parts from bulk bins into
Open source S2
which translates visual information and instructions into physical robot movements. It is currently rolling out to Google’s robotics partners. Model tackles 200 tasks Google evaluated Gemini Robotics 2.5 across more than 200 distinct tasks, including serving food with tongs, tying bows on gift boxes, and transferring liquids with pipettes. These activities test a range of capabilities, from handling delicate objects
Open source S4
to manipulating flexible materials and using everyday tools. Compared with previous baselines, the model achieved success rates that were 30 percent to 100 percent higher, depending on the task. It also completed activities 20 percent to 30 percent faster, according to Google. The model works across different robotic platforms, including Apptronik’s Apollo 2 humanoid and the Franka Duo system, which uses two-fingered
Open source S5
navigate physical hazards. Its inference-time reasoning guidance helps robots avoid obstacles and operate around people. In company evaluations, the capability increased safe task completion by 33 percent around obstacles and 43 percent near people. Google also introduced ASIMOV-Jailbreak, an open-source benchmark designed to test whether robotic systems can resist disguised unsafe instructions. It contains hundreds
Open source S7
ory result remains a proof-of-concept, rather than evidence of widespread commercial deployment. Gemini Robotics 2.5 is currently being rolled out to Google’s robotics partners.
Open source S10
Source：Robotics — Web news discovery · msn.com