DeepMind’s Gemini Robotics 2.5 has a 96% factory claim—and a reliability gap

Counts from X are delayed. They’ll show here once X sends them.

Counts from X are delayed. They’ll show here once X sends them.

There’s also a trade-off between learning many tasks and reliably completing one. Ars quotes Parada: “We can be very good at a particular application at the cost of generalization.” Source

Bows, pipettes and a cloud connection

Gemini Robotics 2.5 is a vision-language-action model: the component that generates a robot’s physical actions. Ars distinguishes it from Gemini Embodied Reasoning, which uses instructions and live video to plan a task. Source

According to Ars, DeepMind evaluated the model across more than 200 activities, including serving food with tongs and using a pipette. The company also showed Apollo 2 tying a bow with its five-fingered hands. Ribbon has apparently joined the robotics curriculum. Source

In DeepMind’s laboratory tests, success rates were 30% to 100% higher than the previous version and execution was 20% to 30% faster, according to Ars. Source

Gemini 4 Argon helped accelerate development by generating simulated practice runs and analyzing failed attempts, according to the team. Ars reports that the team generated thousands of practice-run variations and combined synthetic data with traditional training demonstrations. Source Source

Robotics 2.5 runs on cloud-based servers and has latency, Ars reports. Argon’s failure analysis doesn’t happen in real time, either. Source

A jailbreak benchmark isn’t a safety certificate

DeepMind says new hazard-handling features help robots detect people and obstacles in real time. Ars describes behavior such as stopping or waiting until an obstruction clears. Source Source

Alongside the model, DeepMind introduced an open-source safety benchmark. Ars identifies it as ASIMOV-Jailbreak and describes a database of 1,500 prompts designed to trick robotic AI into unsafe behavior. Source Source

Testing responses to unsafe prompts doesn’t, by itself, establish how safely robots will operate on a factory floor. Source