Embodied Agent Interface (EAI): Benchmarking LLMs for Embodied Decision Making

Manling Li, Shiyu Zhao, Qineng Wang, Kangrui Wang, Yu Zhou, Sanjana Srivastava, Cem Gokmen, Tony Lee, Li Erran Li, Ruohan Zhang, Weiyu Liu, Percy Liang, Li Fei-Fei, Jiayuan Mao, Jiajun Wu

Stanford Vision and Learning Lab, Stanford University

Dataset Highlights

• Standardized goal specifications.
• Standardized modules and interfaces.
• Broad coverage of evaluation and fine-grained metrics.
• Please find our dataset at this link.
• PDDL files for both BEHAVIOR (domain file, problem files) and VirtualHome (domain file, problem files).

Overview

We aim to evaluate Large Language Models (LLMs) for embodied decision-making. While many works leverage LLMs for decision-making in embodied environments, a systematic understanding of their performance is still lacking. These models are applied in different domains, for various purposes, and with diverse inputs and outputs. Current evaluations tend to rely on final success rates alone, making it difficult to pinpoint where LLMs fall short and how to leverage them effectively in embodied AI systems.

To address this gap, we propose the Embodied Agent Interface (EAI), which unifies:

1. A broad set of embodied decision-making tasks involving both state and temporally extended goals.
2. Four commonly used LLM-based modules: goal interpretation, subgoal decomposition, action sequencing, and transition modeling.
3. Fine-grained evaluation metrics, identifying errors such as hallucinations, affordance issues, and planning mistakes.

Our benchmark provides a comprehensive assessment of LLM performance across different subtasks, identifying their strengths and weaknesses in embodied decision-making contexts.

Installation

1. Create and Activate a Conda Environment:

   conda create -n eai-eval python=3.8 -y 
   conda activate eai-eval

2. Install eai:

   You can install it from pip:

   pip install eai-eval

   Or, install from source:

   git clone https://github.com/embodied-agent-interface/embodied-agent-interface.git
   cd embodied-agent-interface
   pip install -e .

3. (Optional) Test PDDL planner for transition modeling: If you want to evaluate transition_modeling, it is highly recommended to test the installation of PDDL planner. You can test by running

   python examples/pddl_tester.py

   If the output is Results: ['walk_towards character light', 'switch_on character light'], the installation is successful. Otherwise, you can refer to the BUILD.md under pddlgym_planners/ or this for more instructions.

4. (Optional) Install iGibson for behavior evaluation:

   If you need to use behavior_eval, install iGibson. Follow these steps to minimize installation issues:

   • Make sure you are using Python 3.8 and meet the minimum system requirements in the iGibson installation guide.

   • Install CMake using Conda (do not use pip):

     conda install cmake

   • Install iGibson: We provide an installation script:

     python -m behavior_eval.utils.install_igibson_utils

     Alternatively, install it manually:

     git clone https://github.com/embodied-agent-interface/iGibson.git --recursive
     cd iGibson
     pip install -e .

   • Download assets:

     python -m behavior_eval.utils.download_utils

   We have successfully tested installation on Linux, Windows 10+, and macOS.

Quick Start

1. Arguments:

   eai-eval \
     --dataset {virtualhome,behavior} \
     --mode {generate_prompts,evaluate_results} \
     --eval-type {action_sequencing,transition_modeling,goal_interpretation,subgoal_decomposition} \
     --llm-response-path <path_to_responses> \
     --output-dir <output_directory> \
     --num-workers <number_of_workers>

   Run the following command for further information:

   eai-eval --help

2. Examples:

• Evaluate Results

  Make sure to download our results first if you don't want to specify <path_to_responses>

  python -m eai_eval.utils.download_utils

  Then, run the commands below:

  eai-eval --dataset virtualhome --eval-type action_sequencing --mode evaluate_results
  eai-eval --dataset virtualhome --eval-type transition_modeling --mode evaluate_results
  eai-eval --dataset virtualhome --eval-type goal_interpretation --mode evaluate_results
  eai-eval --dataset virtualhome --eval-type subgoal_decomposition --mode evaluate_results
  eai-eval --dataset behavior --eval-type action_sequencing --mode evaluate_results
  eai-eval --dataset behavior --eval-type transition_modeling --mode evaluate_results
  eai-eval --dataset behavior --eval-type goal_interpretation --mode evaluate_results
  eai-eval --dataset behavior --eval-type subgoal_decomposition --mode evaluate_results

• Generate Pormpts

  To generate prompts, you can run:

  eai-eval --dataset virtualhome --eval-type action_sequencing --mode generate_prompts
  eai-eval --dataset virtualhome --eval-type transition_modeling --mode generate_prompts
  eai-eval --dataset virtualhome --eval-type goal_interpretation --mode generate_prompts
  eai-eval --dataset virtualhome --eval-type subgoal_decomposition --mode generate_prompts
  eai-eval --dataset behavior --eval-type action_sequencing --mode generate_prompts
  eai-eval --dataset behavior --eval-type transition_modeling --mode generate_prompts
  eai-eval --dataset behavior --eval-type goal_interpretation --mode generate_prompts
  eai-eval --dataset behavior --eval-type subgoal_decomposition --mode generate_prompts

• Simulation

  To see the effect of our magic actions, refer to this notebook.

3. Evaluate All Modules in One Command

   To evaluate all modules with default parameters, use the command below:

   eai-eval --all

   This command will automatically traverse all unspecified parameter options.

   Example Usage:

   eai-eval --all --dataset virtualhome

   This will run both generate_prompts and evaluate_results for all modules in the virtualhome dataset. Make sure to download our results first if you don't want to specify <path_to_responses>

Docker

We provide a ready-to-use Docker image for easy installation and usage.

First, pull the Docker image from Docker Hub:

docker pull jameskrw/eai-eval

Next, run the Docker container interactively:

docker run -it jameskrw/eai-eval

Test docker

eai-eval

By default, this will start generating prompts for goal interpretation in Behavior.

BibTex

If you find our work helpful, please consider citing it:

@inproceedings{li2024embodied,
  title={Embodied Agent Interface: Benchmarking LLMs for Embodied Decision Making},
  author={Li, Manling and Zhao, Shiyu and Wang, Qineng and Wang, Kangrui and Zhou, Yu and Srivastava, Sanjana and Gokmen, Cem and Lee, Tony and Li, Li Erran and Zhang, Ruohan and others},
  booktitle={NeurIPS 2024},
  year={2024}
}