HomeFlow: A Data Flywheel for Smart Home Agent Training with Verifiable Simulation

· Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

HomeFlow is a novel data flywheel designed to generate high-quality training data for large language model agents controlling smart home environments. This system addresses challenges in real domestic interaction, such as ambiguous intents and dynamic environments. HomeFlow integrates HomeEnv as a unified simulation environment and HomeMaker for procedurally generating diverse smart home settings. It employs Blueprint to compile open-ended user intents into executable state-based success conditions, while MCTS-Flow synthesizes verifiable multi-turn trajectories through environment-guided tree search. Agents are then optimized using supervised fine-tuning and step-wise RLVE, facilitating iterative improvement. Evaluated on the SmartHome-Bench, HomeFlow-RL-4B and HomeFlow-RL-8B achieved task success rates of 84.60% and 87.03% respectively, with HomeFlow-RL-8B notably outperforming GPT-5.5 by 1.23 percentage points.

Key takeaway

For AI Engineers developing LLM agents for smart home control, HomeFlow demonstrates a robust approach to generating high-quality training data. You should consider integrating verifiable simulation environments and procedural generation techniques to overcome challenges with ambiguous intents and dynamic environments. This method, which includes environment-guided tree search and iterative RLVE, can significantly improve agent task success rates, potentially surpassing current leading models like GPT-5.5.

Key insights

HomeFlow uses a verifiable data flywheel with simulation and search to train smart home agents effectively.

Principles

Method

HomeFlow's method involves using HomeEnv and HomeMaker for environment setup, Blueprint for intent compilation, MCTS-Flow for trajectory synthesis, and supervised fine-tuning with step-wise RLVE for agent optimization.

In practice

Topics

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.