Google Unveils Gemini Robotics 2.5 and Gemini 4 Argon for Physical Intelligence
New models enhance robot dexterity and reasoning for complex industrial and simulation-based tasks.
1Advancing Physical Intelligence
Google introduced Gemini Robotics 2.5, a vision-language-action model designed to improve robot dexterity and reliability. Simultaneously, the company released Gemini 4 Argon, a general-purpose physical-reasoning model. These systems aim to transition robots from controlled laboratory settings into unpredictable, unstructured real-world environments, enabling them to handle diverse tasks with human-level precision.
2Enhanced Task Performance
Gemini Robotics 2.5 was evaluated across 200 distinct tasks, including pipetting liquids and tying bows. Compared to previous baselines, the model achieved success rate increases ranging from 30% to 100%. Furthermore, the system completed these complex operations 20% to 30% faster, demonstrating significant improvements in speed, robustness, and overall task precision.
3Industrial Pilot Success
In partnership with Mercedes-Benz and Apptronik, Gemini Robotics 2.5 was deployed on Apollo 2 humanoid robots in Hungary and Germany. The robots autonomously performed automotive kitting, sorting parts from bulk bins into trays. The pilot achieved a 96% success rate, proving the model's capability to function effectively within active production lines.
4Safety and Navigation
To operate safely alongside humans, Gemini Robotics 2.5 utilizes inference-time reasoning guidance. This feature allows robots to identify and steer around physical hazards in real-time. Evaluations showed this capability boosted safe task completion around obstacles by 33% and improved safety when working in close proximity to people by 43%.
5Gemini 4 Argon Capabilities
Gemini 4 Argon functions as a physical-reasoning model for robot control, scene construction, and iterative policy improvement. It scored 87% on the Embodied Reasoning Question Answering benchmark using pure reasoning without helper tools. The model is currently available to Google AI Ultra subscribers and through the paid Gemini API for developers.
6Simulation and Real-World Links
Argon enables Real2Sim scene construction, allowing developers to rebuild MuJoCo simulation environments from teleoperation video. By synthesizing domain-randomized episodes from single RGB photos, Google reported a success rate increase from 56% to 75% on real-world setups. This workflow helps fine-tune vision-language-action policies by bridging the gap between simulation and physical reality.
7Technical Implementation Details
The robotics system utilizes long, task-specific prompts, coordinate grids, and motion primitives to guide robot behavior. While the system demonstrates advanced reasoning, some developers note that simulation fidelity remains critical. Matching contact events is considered as important as visual accuracy, as scenes can appear convincing while objects behave incorrectly during manipulation.
8Future of Embodied AI
These models represent a shift toward integrating reasoning into physical control loops. By combining Gemini Robotics 2.5 for action and Gemini Robotics ER for spatial planning, Google aims to scale humanoid deployment. The industry now faces the question of whether these frontier models will replace traditional pipelines of separate simulation services.