HomeFlow: A Data Flywheel for Smart Home Agent Training with Verifiable Simulation
Researchers have introduced HomeFlow, a novel data flywheel designed to improve the training of AI agents for smart home environments. This system utilizes a unified simulation environment and procedural generation of home settings to create diverse and verifiable training data. HomeFlow synthesizes multi-turn trajectories and optimizes agents through fine-tuning and reinforcement learning, achieving high task success rates on a new benchmark and even outperforming GPT-5.5. AI
IMPACT This research could accelerate the development of more capable AI agents for controlling smart home devices and other real-world applications.