On 20 November 2026, Google DeepMind announced the availability of new versions of the Gemini 4 and Genie 4 models, which substantially expand the capabilities of generating realistic world simulations and their application in robot training. These updates build on previous work in the field of world models and vision-language-action systems, but introduce key improvements in real time and multimodality.
Genie 4 is a natively multimodal world model that generates not only visual scenes at resolutions up to 720p at 24 fps, but also synchronized audio content: ambient sounds, music, and character dialogue. The model integrates the capabilities of Gemini Omni for anchoring simulations in the real world and other Gemini variants to maintain coherence and continuity of scenes. Users can influence the simulation in real time through text prompts — changing the weather, introducing objects or characters — while the Gemini components adapt the changes contextually and monitor physical plausibility.
In the field of robotics, a key role is played by Gemini 4 Argon, which acts as a simulation engine. It generates thousands of varied simulated runs based on real video scenes, creating digital twins and accelerating skill generalization. This made it possible to develop Gemini Robotics 2.5 — the most advanced VLA model to date. According to DeepMind, the model demonstrates an increase in task success rate of 30–100 % depending on the type and an acceleration of execution of 20–30 % compared with previous versions, tested on more than 200 tasks, including work with deformable objects and delicate tools.
The evaluation methodology includes both laboratory tests and real-world pilots, for example on Mercedes-Benz production lines with the Apptronik Apollo 2 robot, where a 96 % success rate was achieved in the parts kitting task. However, DeepMind does not publish detailed benchmarks compared with external models, and the focus is placed on internal metrics and steerability — the ability to avoid hazards in real time. This leaves questions about reproducibility and comparison with parallel approaches from other laboratories, such as the work of OpenAI or Boston Dynamics in embodied AI.
In the landscape of world models, Genie 4 builds on Genie 3, doubling the duration of coherent simulations and adding native audio, which distinguishes DeepMind's approach from purely visual generators such as Runway. The integration with Gemini 4 Argon for robotics creates a closed loop: simulations accelerate the training of VLA models, and these in turn improve the quality of simulations. This is a strategic shift compared with previous iterations, where data was collected predominantly in the real world.
The downstream consequences include an acceleration of the development of physical intelligence: robots adapt faster to new environments, which opens the path to industrial scaling. However, it remains unclear how generalizable the skills are beyond the tested 200 tasks and how the model copes with long-term agentic scenarios without additional checks.
Open questions concern independent verification of the claimed improvements and the impact on safety — DeepMind released the ASIMOV-Jailbreak benchmark, where Argon shows 99,5 % robustness, but real risks in uncontrolled environments require further research by the community.
Thus, the combination of Gemini 4 and Genie 4 shifts the emphasis from isolated models to integrated simulation-training loops, which makes the next step in embodied AI more attainable.

