Google reports Gemini Robotics 2.5 reached 96% factory task success

• Gemini Robotics 2.5 is rolling out to robotics partners; Apollo 2 reached 96% task success in Google’s Mercedes-Benz proof of concept.
• Across 200 tasks, success rates rose 30% to 100% and execution was 20% to 30% faster than previous baselines, according to Google’s evaluation results.
• Gemini 4 Argon helped generate simulated practice and diagnose failures during development, as Google described in its launch thread.
• Safety additions cover physical hazard avoidance and adversarial instructions, according to Google DeepMind’s safety announcement.

Apollo 2’s five-finger SharpaWave hand has 22 degrees of freedom, while the same model also works with two-finger Franka Duo grippers. Apptronik co-founder and CEO Jeff Cardenas says a few weeks of post-training transferred to Mercedes production with “very little drop-off.”

Google tested Apollo 2 on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The task was automotive kitting: moving metal brackets from bulk supply bins into sequencing trays, ordered for each vehicle’s assembly workflow.

Apollo 2 used a two-model stack, the pilot’s most revealing engineering detail:

• Gemini Robotics ER: tracked the workflow and reasoned over multiple steps.
• Gemini Robotics 2.5: controlled the hands and arms, picking and placing parts while correcting gaps.

The announcement reports 96% success but does not disclose trial counts, operating hours, intervention rates or cycle times.

Google evaluated 200 distinct tasks in its capability tests, including:

• Tying a bow: manipulating deformable objects.
• Pipetting liquid: handling delicate objects precisely.
• Serving food with tongs: using an everyday tool.

Against what it calls “previous baselines,” Google reports:

• Task success: 30% to 100% higher, depending on task type.
• Execution speed: tasks completed 20% to 30% faster.

Those are relative improvements; Google does not identify the comparison baselines or give per-task absolute scores in the launch post.

Google describes a data-generation and debugging process in its technical announcement:

1. Generate practice: Argon builds a realistic digital twin from a real-world camera scene and generates thousands of varied simulated runs.
2. Combine training sources: simulated practice joins real demonstrations to improve generalization to new scenes and objects.
3. Diagnose failures: Argon analyzes video footage, helps researchers hypothesize causes, and supports testing improvements in simulation and on physical robots.

Gemini Robotics 2.5 uses reasoning guidance during inference to steer around physical hazards. In Google’s evaluations, that guidance improved safe task completion:

• Around obstacles: 33% higher.
• Around nearby people: 43% higher.

These figures measure successful task completion under safety constraints, rather than reductions in collision rates, as described in the safety results.

Google reports 99.5% jailbreak resistance for Gemini 4 Argon, versus 82% for Gemini Robotics ER 2, a 17.5-percentage-point increase in its launch evaluation. The reported score belongs to Argon, the reasoning model, rather than the Robotics 2.5 actions model.

The released evaluation harness exposes more detail than the headline score:

• 845 multi-turn jailbreak episodes across 15 attack-strategy groups, with up to 10 steps per episode.
• Simulated read-only inspection and calibration tools available during those episodes.
• Paired harmful and benign tabletop tasks, testing implicit safety behavior and an explicit safety policy.
• Deterministic action-tool checks plus an LLM judge, with attack-success and benign over-refusal metrics and a 41-criterion physical-safety rubric.

The harness defaults to Argon as its LLM judge and supports scoring with multiple judges simultaneously.