Embodied Simulation and Transfer Evidence Record

Embodied Simulation and Transfer Evidence Record

Development value

Virtual environments make some embodied-agent experiments cheaper, faster, repeatable, parallel, and safer to fail. A team can vary scenes, reproduce a failure, collect interaction data, and compare policies without consuming physical hardware for every attempt.

SIMA used commercial games and research environments as interactive 3D worlds. Its interface accepted image observations and language instructions and produced keyboard-and-mouse actions. This made the control surface broadly reusable across environments.

Figure's 2026 Helix 02 record describes training a whole-body controller in more than 200,000 parallel simulated environments with domain randomization before direct transfer to real robots. That is a first-party example of simulation inside a later physical-robot workflow. It is not evidence that every simulated behavior transfers.

Transfer boundary

Every simulator contains assumptions. An agent can exploit shortcuts that do not exist outside the virtual world. Even a successful transfer can remain narrow to the tested robot, task, environment, and configuration.

Current SIMA chronology

Google DeepMind introduced SIMA in March 2024. The technical report described preliminary results across research environments and commercial games. In November 2025, Google DeepMind introduced SIMA 2, a Gemini-based follow-on that added dialogue, more complex instruction handling, and a self-improvement experiment in virtual worlds.

SIMA 2 remains a virtual-world research record. Google DeepMind describes eventual physical-world implications as a direction, not as completed robot deployment evidence.

Publication use

The essay should present simulation as a learning-loop accelerator. It should avoid claiming that game agents and physical robots are the same system, that SIMA controls Figure robots, or that simulation certifies physical safety.