Gemini Robotics 2.5: humanoids pass a Mercedes plant pilot

What changed

Gemini Robotics 2.5 is DeepMind's VLA model: it turns camera input and a text instruction directly into control of a robot's hands and arms. Planning still sits with Gemini Robotics ER, while 2.5 produces the actions (Google DeepMind).

• Evaluated on more than 200 distinct tasks, such as serving food with tongs, tying a bow on a gift box and pipetting.
• Compared with the previous baseline, success rates are 30% to 100% higher depending on task type, and tasks finish 20% to 30% faster.
• Works across robot types, including Apptronik's Apollo 2 humanoid (five-finger, 22-degree-of-freedom SharpaWave hand) and the dual-arm Franka Duo.
• It is rolling out to robotics partners. There is no public API or downloadable weights.

On the factory floor

The real-world test ran on Mercedes-Benz production lines in Kecskemét, Hungary and Sindelfingen, Germany. Apollo 2 performed kitting: sorting metal brackets from bulk bins into sequenced trays in the exact order each vehicle on the line needs. The pilot reached a 96% success rate.

Apptronik's CEO says the models improved sharply after a few weeks of post-training and carried over to Mercedes-Benz production with very little drop-off. It is a proof of concept, not a rollout, but production-line numbers are rare in robotics announcements.

Safety and a new jailbreak benchmark

Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around physical hazards. DeepMind says this raised safe task completion by 33% around obstacles and by 43% around nearby people.

DeepMind also released ASIMOV-Jailbreak, an open benchmark with hundreds of prompts that disguise unsafe requests across dozens of scenes. Gemini 4 Argon scored 99.5% jailbreak resistance, up from 82% for the previous baseline, Gemini Robotics ER 2 (Gemini 4 Argon deep dive: Google's new frontier model).

Why it matters

Argon also acted as a simulation engine, generating thousands of simulated practice runs and a digital twin from a single real camera scene, which sped up training. It has been a busy week in robotics, with NVIDIA also showing a new GR00T model (GR00T 2: 71% robot task success from one demo). The takeaway for practitioners: humanoid pilots are now producing measurable results on real automotive lines, with a 96% success rate as the bar to beat.