SmallPlan: Leverage Small Language Models for Sequential Path Planning with Simulation-Powered, LLM-Guided Distillation

Updating...

Code overview

SmallPlan adopts the open-sourced code from MoMaLLM [paper] and TextGames [paper]

• We use MoMaLLM code base for simulation and LLM agent set-up. We then add functions for fine-tuning the SLMs with SFT and RL.

• We use TextGames for out-of-domain analysis experiment. We mostly run ./textgames/agents/dsr1_distill.py to test our fine-tuned SLMs.

The weight of the fine-tunned SLMs can be found here

Our code requires 2 different python venv to run. This is due to the module conflicts of igibson environment versus the unsloth environment for training. We use the smallplan environment to set up the SLMTrainer and expose it externally through FastAPI. The igibson environment is then used to run the simulation, which communicates with the SLMTrainer by making API calls for each training epoch.

Alternatively, the SLMTrainer API could be hosted in a Docker container. However, we don't do this because our server is already Docker-based, and it does not permit creating child containers within the main container.

Local installation

iGibson prerequisite

1. Install iGibson: https://stanfordvl.github.io/iGibson/installation.html
2. Update the conda environment for igibson with the dependencies with pip3 install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu113

Python venv setup

For smallplan environment (Python 3.11), pip install -r requirements.api.txt is all you need.

For igibson environment (Python 3.8), pip install -r requirements.txt is all you need.

Please run cp .env.example .env in the terminal and add your WANDB_API_KEY, OPENAI_ORGANIZATION, OPENAI_API_KEY

Training

First, please go to ./configs/slm_training.yaml and ./configs/wandb.yaml and chaneg your config.

Open a terminal and run

conda activate smallplan
(smallplan) python -m src.slm_api.app

Open another terminal and run

conda activate igibson
(igibson) python -m src.train_from_simulation.app

Citation

Please cite our work if you find helpful.

@misc{pham2025smallplanleveragesmalllanguage,
      title={SmallPlan: Leverage Small Language Models for Sequential Path Planning with Simulation-Powered, LLM-Guided Distillation}, 
      author={Quang P. M. Pham and Khoi T. N. Nguyen and Nhi H. Doan and Cuong A. Pham and Kentaro Inui and Dezhen Song},
      year={2025},
      eprint={2505.00831},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2505.00831}, 
}