SIMA Research Design and Follow-On Record

SIMA Research Design and Follow-On Record

SIMA 1 question

Google DeepMind's 2024 SIMA work asked whether one agent could follow free-form language instructions across varied simulated 3D environments. The system used image observations and language as input and produced keyboard-and-mouse actions through a broadly shared interface.

The researchers used curated environments and commercial games. The goal was language-driven generality across environments rather than maximizing a game score.

Research evidence

The technical report presents preliminary results, environment and task design, data collection, evaluation, and limitations. Google DeepMind's announcement says the initial agent learned more than 600 language-following skills. That is a first-party research summary. Exact benchmark claims should be taken from the technical report with the named split, environment, baseline, and metric.

SIMA did not establish unrestricted autonomy, consciousness, a self-generated purpose, physical-robot control, or deployment in an uncontrolled world.

Follow-on

Google DeepMind announced SIMA 2 in November 2025. It integrated Gemini capabilities, accepted language and image instructions, added dialogue and higher-level reasoning, and included self-improvement experiments inside virtual environments.

The SIMA 2 technical report reports improved performance and generalization across training and held-out virtual environments. It remains research in virtual worlds. The authors describe physical-world relevance as an eventual direction.

Research-profile boundary

The public profile should preserve four separate ideas: the research goal, the interface, the evaluated environments, and the result limits. It should not merge SIMA with NVIDIA simulation tools, Figure robots, or a public product.

The page may explain that games are controllable embodied environments. It must also state that virtual success does not supply a physical safety case, maintenance plan, privacy model, human-intervention record, or real-world reliability result.