Google introduces Gemini Robotics 2.5 with hazard detection and human avoidance

• Gemini Robotics 2.5 adds inference-time hazard avoidance, with Google DeepMind reporting better safe task completion around obstacles and people.
• Testing covered more than 200 tasks, with 20–30% faster completion than previous baselines and a partner rollout underway, according to WesRoth's report.
• The 99.5% jailbreak-resistance result belongs to Gemini 4 Argon, as one launch breakdown noted, rather than the Robotics 2.5 actions model.

Argon builds digital twins from a camera scene, then helps researchers diagnose failed robot runs. The open safety harness is the most reusable artifact here: it scores benign over-refusal alongside successful attacks.

Introduced on November 10, Gemini Robotics 2.5 is a vision-language-action model that uses inference-time reasoning guidance to steer around physical hazards. Google reports improvements in safe task completion of:

• 33% around obstacles.
• 43% around nearby people.

These metrics measure improvements in completing tasks safely; the announcement does not provide absolute collision rates or the guidance's inference overhead.

Google evaluated more than 200 tasks and reported 20–30% faster completion, as WesRoth noted. Its launch report also gives relative task-success gains of 30–100%, depending on task type, against unspecified previous baselines.

The demonstrations span three manipulation problems:

• Deformable objects: tying bows and folding clothes.
• Delicate handling: pipetting liquids.
• Tool use: serving pastries with tongs.

Supported hardware includes Apollo 2's five-finger, 22-degree-of-freedom SharpaWave hand and Franka Duo's two-finger parallel grippers.

Bow tying and delicate-object manipulation

Google describes a simulation-and-analysis loop for accelerating robotics development:

1. Generate practice: Argon creates thousands of varied simulated runs from a real-world camera scene.
2. Mix training data: simulated practice combines with real demonstrations to help robots generalize to new scenes and objects.
3. Test corrections: researchers use Argon's hypotheses about failures to test improvements in simulation and on physical robots.

Google says Argon helped accelerate Robotics 2.5's development; it gives no numerical training-time reduction in the announcement.

Apollo 2 achieved 96% success in a proof of concept on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Its job was automotive kitting: moving metal brackets from bulk bins into containers sequenced for vehicle assembly.

The pilot split responsibilities between two models:

• Gemini Robotics ER: tracked the workflow.
• Gemini Robotics 2.5: controlled the robot's hands and arms.

Argon reached 99.5% jailbreak resistance, as one launch breakdown highlighted, versus Google's 82% baseline for Gemini Robotics ER 2, a 17.5-percentage-point gain.

The released evaluation harness exposes two complementary suites:

1. Paired harmful and benign tabletop tasks: compares tool declarations alone with an explicit safety policy.
2. Multi-turn jailbreaks: 845 episodes across 15 attack-strategy groups, with up to 10 steps and simulated read-only inspection and calibration tools.

Scoring combines deterministic action-tool checks with an LLM judge. Argon is also the default judge, although the harness supports multiple judges simultaneously.

In Robocurve's separate study, five multimodal language models, including Argon, controlled physical robot arms across five unsafe tasks. Across 300 trials per condition, unsafe-attempt rates were:

• Unaltered video: 27%.
• Gaussian-blurred video: 88%.
• Doctored video: 97%.

The experiment used 12 trials per model, task and condition, with a language model classifying attempts versus refusals. It tested general-purpose multimodal models through a robot-control harness; Gemini Robotics 2.5 was outside its test set.