Google DeepMind has introduced the Gemini-powered SIMA 2. The agent, which understands and acts in virtual worlds, continuously improves its performance by learning from its own experiences.
Google DeepMind shared a comprehensive research preview on Thursday for SIMA 2, which represents the next major step in artificial intelligence research. The new-generation general-purpose agent, combined with Gemini’s advanced language and reasoning capabilities, is no longer just a system that follows commands. Instead, it gains the ability to understand and interact with the virtual world it inhabes.
SIMA 2 Raises the Bar
DeepMind had trained the first version, SIMA 1, with hundreds of hours of game footage, demonstrating that the agent could play numerous 3D games like humans. However, SIMA 1’s success rate at completing complex tasks was only 31 percent. Humans achieved 71 percent success on the same tasks. It is stated that SIMA 2, developed to overcome these limitations, has both reached a more general level of intelligence and can improve itself by learning from its own experiences.
According to researchers, this feature represents a critical step toward more comprehensive robotics systems and the general-purpose artificial intelligence defined as AGI.
Furthermore, SIMA 2 is powered by the Gemini 2.5 Flash-Lite model. This structure, defined as an Embodied Agent, observes its surroundings by interacting with a physical or virtual world through a “body” and produces actions accordingly. This approach differs from traditional AI, which only performs abstract operations like calendar management or code execution.
Jane Wang, a senior researcher at DeepMind, emphasizes that SIMA 2 is no longer just playing games but can grasp user instructions with their context and provide logical, consistent, and common-sense responses. Thanks to the Gemini integration, SIMA 2’s performance has doubled compared to the previous version.
It Trains and Learns on Its Own
In a live demo on “No Man’s Sky,” SIMA 2 described the planet’s rocky surface, recognized a nearby emergency beacon, and determined its next step. In another example, when given the command “go to the house the color of a ripe tomato,” it thought, “A tomato is red, so I must go to the red house,” and then proceeded to the red house. SIMA 2 can also understand emoji-based commands. For instance, when sent an axe and a tree emoji, the agent interprets this and goes to chop down a tree. The agent can also recognize correct objects and interact with details like benches, trees, and butterflies in new photorealistic worlds created with DeepMind’s Genie model.
One of the most striking innovations is its self-learning capacity. While SIMA 1 was trained entirely on human gameplay data, SIMA 2 only takes its initial foundation from this data. Afterward, when the system is placed in new environments, it generates tasks from another Gemini model, and an independent reward model scores the agent’s performance. In this cycle, SIMA 2 learns from its own mistakes and develops new behaviors with feedback generated entirely by AI.
DeepMind sees SIMA 2 as a gateway to more comprehensive robotics platforms in the future. Researchers state that a robot operating in the real world needs high-level comprehension and reasoning skills, and SIMA 2 operates precisely at this upper layer. In contrast, low-level control mechanisms, such as physical joints or wheels, are managed by different models. For now, there is no set timeline for integrating SIMA 2 into physical robots or releasing a public version.
You Might Also Like;
