Simulating the Resident: Generating Executable Smart Home Schedules via LLM Personas

Technical Summary: Simulating the Resident via LLM Personas

Problem Statement
Research in Human-Computer Interaction (HCI), security, and privacy for smart homes relies heavily on datasets capturing authentic device interactions, network traffic, and daily routines. However, acquiring such data presents significant methodological challenges. Collecting data from real households is slow, expensive, and raises profound privacy concerns due to the need for long-term, invasive observation of private spaces. Furthermore, existing real-world datasets are often limited in scope, failing to generalize across diverse household compositions, device ecosystems, or usage patterns. While synthetic data generation has been explored, previous approaches (e.g., statistical models or Hidden Markov Models) lack a semantic model of human intent, and existing LLM-based simulations (e.g., AgentSense) remain confined to virtual environments without connecting to real hardware.

Methodology
The authors propose a novel approach to simulate residents themselves using Large Language Models (LLMs) to generate diverse, behaviorally grounded interaction schedules. These schedules are designed to be executable on physical smart-home testbeds. The methodology consists of two integrated components:

1. A Design Framework for Households:
   The framework configures simulated households across five socio-technical dimensions to shape device activity:

   • Occupational Routines: Work patterns (e.g., 9-to-5, shift work) that structure daily life and presence.
   • Simulation Timeframe: The temporal scope (e.g., single-day snapshot, weekly, longitudinal).
   • Household Dynamics: Social structures and coordination (e.g., single resident, cooperative cohabitants, conflicting schedules).
   • Device Ecosystem & Interaction Style: The installed hardware and technical literacy, determining automation complexity.
   • Environmental Context: Physical settings and climate (e.g., urban apartment, winter conditions) influencing spatial and triggered behaviors.

2. A Multi-Stage LLM Pipeline:
   Adapted from the generative agent architecture, the pipeline translates the configured dimensions into executable traces through four stages:

   • Stage 1: Persona & Context Initialization: The LLM receives the household configuration, environmental context, and a strict JSON schema of available devices. It generates a "persona memory card" for each resident, summarizing their schedule and habits while explicitly mapping preferences to exact device names and parameter values from the schema.
   • Stage 2: Narrative Day-Plan Generation: Using the memory cards, the LLM generates a natural-language narrative of activities within a discrete time window. This stage emphasizes reasoning about physical locations and social coordination (e.g., avoiding waking another resident). Critical constraints enforce exact timestamps, explicit state management, definitive language, and strict adherence to the device schema.
   • Stage 3: Action Extraction and Formatting: A strict formatting prompt instructs the LLM to act as an action parser, extracting concrete device interactions from the narrative into a structured JSON array. This step separates reasoning from formatting to reduce syntax errors and hallucinations. The output includes timestamps, resident names, devices, actions, values, and intents.
   • Stage 4: State and Memory Update: For longitudinal simulations, the pipeline updates persona memory cards based on the generated actions. This rolling memory mechanism ensures spatial and behavioral consistency across extended simulations without exceeding context windows.

Key Contributions
The paper presents three primary contributions:

1. A Configurable Design Framework: A set of five socio-technical dimensions that allow researchers to systematically shape simulated household scenarios.
2. A Multi-Stage LLM Pipeline: A detailed architecture that transforms high-level resident personas and context into structured, executable device interaction schedules.
3. A Proof of Concept: An initial demonstration validating that the pipeline can produce temporally coherent, persona-consistent, and schema-compliant interaction traces.

Results
In a proof-of-concept experiment simulating two cohabitants (a work-from-home professional and an office worker) in a German winter morning, the pipeline successfully generated a morning routine (06:00–10:00) for eight smart devices.

• Contextual Coherence: The outputs reflected the seasonal context (e.g., increased heating usage due to winter sunrise) and distinct temporal patterns for each resident (e.g., Bob's actions clustered before departure, Alice's spread throughout the morning).
• Schema Compliance: The generated JSON schedules strictly adhered to the provided device schema, correctly separating actions from parameters and utilizing only defined device names.
• Feasibility: The pipeline successfully translated high-level traits and environmental constraints into a format ready for execution on physical testbeds (e.g., Home Assistant, TTDAS).

Significance and Claims
The authors position this work as a "work in progress" aimed at supporting scalable, privacy-conscious smart-home experimentation without relying on intrusive real-world data collection. They claim the approach offers three specific implications:

• Direct Testbed Execution: The ability to drive physical hardware with generated routines allows researchers to capture authentic network traffic and device states.
• Standardized Benchmarks: The socio-technical dimensions facilitate the creation of shareable, off-the-shelf household configurations for evaluating systems against baseline behaviors.
• Rapid Prototyping: The simulation serves as a tool for designers to preemptively test multi-user conflicts and automation logic.

The paper modestly acknowledges that ecological validity has not yet been established, particularly regarding the ability of LLMs to capture the irregular or forgetful nature of real domestic behavior. Furthermore, end-to-end execution on a physical testbed has not yet been demonstrated in this initial study. Future work is directed toward systematic evaluation across diverse configurations, connecting to live testbeds, and potentially building a large-scale open-source dataset of simulated behaviors.