Yes, Genie 3 is specifically designed to support embodied agent research and training. To test the compatibility of Genie 3 created worlds for future agent training, we generated worlds for a recent version of our SIMA agent, our generalist agent for 3D virtual settings. The system has already been tested with DeepMind’s SIMA (Scalable Instructable Multiworld Agent), demonstrating practical compatibility with existing agent architectures.
The training process works by having agents interact with Genie 3 environments through standard action interfaces. In each world we instructed the agent to pursue a set of distinct goals, which it aims to achieve by sending navigation actions to Genie 3. Like any other environment, Genie 3 is not aware of the agent’s goal, instead it simulates the future based on the agent’s actions. This approach allows agents to learn from their interactions without the world model being biased toward specific objectives, creating more robust training scenarios where agents must develop genuine understanding and problem-solving capabilities.
The extended consistency that Genie 3 provides is particularly valuable for agent training. Since Genie 3 is able to maintain consistency, it is now possible to execute a longer sequence of actions, achieving more complex goals. This enables training on multi-step tasks that require planning, memory, and sequential decision-making. The researchers explicitly state that world models are also a key stepping stone on the path to AGI, since they make it possible to train AI agents in an unlimited curriculum of rich simulation environments. The system’s ability to generate diverse environments from text prompts means researchers can create unlimited training scenarios, from realistic environments for robotics to fantastical settings for testing agent adaptability and generalization capabilities.