Simulating Human-like Daily Activities with Desire-driven Autonomy

Yiding Wang*, Yuxuan Chen*, Fangwei Zhong# , Long Ma,  Yizhou Wang 

Peking University,  The University of Hong Kong,  Beijing Normal University,  BIGAI

* equal contribution    # corresponding author 

Arxiv  |  Github 

Abstract

Existing task-oriented AI agents often depend on explicit instructions or external rewards, limiting their ability to be driven by intrinsic motivations like humans. In this paper, we present a desire-driven autonomy framework to guide a Large Language Model-based ~(LLM-based) agent to simulate human-like daily activities. In contrast to previous agents, our Desire-driven Autonomous Agent (D2A) operates on the principle of intrinsic desire, allowing it to propose and select tasks that fulfill its motivational framework autonomously. Inspired by the Theory of Needs, the motivational framework incorporates an understanding of human-like desires, such as the need for social interaction, personal fulfillment, and self-care. Utilizing a desire-driven task generation mechanism, the agent evaluates its current state and takes a sequence of activities aligned with its intrinsic motivations. Through simulations, we demonstrate that our Desire-driven Autonomous Agent (D2A) generates coherent, contextually relevant daily activities while exhibiting variability and adaptability similar to human behavior. A comparative analysis with other LLM-based frameworks demonstrates that our approach significantly enhances the rationality of the simulated activities. 

Motivation

Our Desire-driven Autonomous Agent (D2A) follows the "Act to Satisfy Intrinsic Desires" principle (right), whereas most existing methods or frameworks focus on goal reasoning or characteristic-driven actions.

The Desire-driven Autonomous Agent Framework

The Desire-Driven Autonomy framework. The red blocks represent procedures from the Value System, the blue blocks indicate procedures from the Desire-Driven Planner, the green blocks highlight the characteristics profile, and the yellow blocks correspond to elements related to the environment controller.

Results

Human-like Results (Win Rate) Evaluated By GPT4o

The win rate among 4 agents. Each point represents the win rate of the agent on the vertical axis when compared with the agent on the horizontal axis. The diagonal values are set to 0.5.

The Averaged Dissatisfaction Results

D2A significantly outperformed the three baselines, exhibiting lower dissatisfaction means and standard deviations.  This result demonstrates that D2A is capable of recognizing both its current desire states and expected states, and can rationally select activities to fulfill its desires, resembling human behavior.

Acknowledgements

This work was supported by the National Science and Technology Major Project (2022ZD0114904), NSFC-6247070125, NSFC-62406010, the State Key Lab of General Artificial Intelligence at Peking University, Qualcomm University Research Grant, and Wuhan East Lake High-Tech Development Zone, National Comprehensive Experimental Base for Governance of Intelligent Society. We thank Wei Wang, Junqi Wang, Prof. Jiayu Zhan, and Prof. Song-Chun Zhu for their helpful discussion in our early work.