D2A

Simulating Human-like Daily Activities with Desire-driven Autonomy

paper link:

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

Desire-driven Autonomous Agent Framework

Environment

The D2A environment is under the examples/D2A folder.

In D2A folder the file structure is:

examples/
├── D2A/
│   ├── Baseline_agent/
│   │   ├── BabyAGI_ActComp.py
│   │   ├── Baseline_BabyAGI.py
│   │   ├── Baseline_comp.py
│   │   ├── Baseline_LLMob.py
│   │   ├── Baseline_ReAct.py
│   │   ├── LLMob_ActComp.py
│   │   └── ReAct_ActComp.py
│   └── D2A_agent/
│       ├── Value_ActComp.py
│       └── ValueAgent.py
├── Environment_construction/
│   └── generate_indoor_situation.py
├── NPC_agent/
│   └── generic_support_agent.py
├── result_folder/
├── value_components/
│   ├── hardcoded_value_state.py
│   └── init_value_info_social.py
│
├── value_comp.py
├── __init__.py
├── experiment_setup_indoor.py
├── experiment_setup_outdoor.py
├── indoor_Room.py
├── NullObservation.py
└── outdoor_party.py

Usage

Run the simulation:

1. Set up the environment dependencies by: ``

2. Set up the experiment setting in experiment_setup_indoor.py and experiment_setup_outdoor.py

   • ROOT (the root path of the project, the folder which has concordia and examples in it)
   • episode_length
   • disable_language_model
   • st_model
   • tested_agents
   • whether to use previous profile
     • Use_Previous_profile
     • previous_profile_file
     • previous_profile
     • previous_profile's path
   • Backbone LLM (details refers to concordia package)
     • api_type
     • model_name
     • api_key
   • desires-related stuff (the desire dimensions to be used while simulating)
     • visible_desires (will appear in the context as the hint for the agent)
     • hidden_desires (will not appear in the context as the hint for the agent)
   • the path to store the result

3. run the simulation by python PATH/to/indoor_Room.py or python PATH/to/outdoor_party.py

Modify desire-related components

Reach the value_components folder and follow these steps across different files:

value_comp.py

1. create the new desire component subclass in value_comp.py

init_value_info_social.py

2. Modify the profile_dict by adding a new entry:

    {"descriptive_adjective": "desire_name"}

3. Add the desire name to the values_names list
4. Add corresponding descriptions to values_names_descriptions

hardcoded_value_state.py

1. Add numerical value mappings to hardcore_state. You can use GPT to design appropriate mapping scales.

Optional: Enable LLM Value Conversion

To use LLM for action value conversion, reach value_comp.py

• replace _convert_numeric_desire_to_qualitative_by_hard_coding(self) function in the _make_pre_act_value(self) with _convert_numeric_desire_to_qualitative(self).