PSEC

Skill Expansion and Composition in Parameter Space

International Conference on Learning Representation (ICLR), 2025

🔥 Official Implementation

PSEC is a novel framework designed to:

🚀 Facilitate efficient and flexible skill expansion and composition
🔄 Iteratively evolve the agents' capabilities
⚡ Efficiently address new challenges

Quick start

Clone this repository and navigate to PSEC folder

git clone https://github.com/ltlhuuu/PSEC.git
cd PSEC

Environment Installation

Environment configuration and dependencies are available in environment.yaml and requirements.txt.

Create conda environment for this experiments

conda create -n PSEC python=3.9
conda activate PSEC

Then install the remaining requirements (with MuJoCo already downloaded, if not see here):

pip install -r requirements.txt

Install the MetaDrive environment via

pip install git+https://github.com/HenryLHH/metadrive_clean.git@main

MuJoCo installation

Download MuJoCo:

mkdir ~/.mujoco
cd ~/.mujoco
wget https://github.com/google-deepmind/mujoco/releases/download/2.1.0/mujoco210-linux-x86_64.tar.gz
tar -zxvf mujoco210-linux-x86_64.tar.gz
cd mujoco210
wget https://www.roboti.us/file/mjkey.txt

Then add the following line to .bashrc:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:~/.mujoco/mujoco210/bin

Run experiments

Pretrain

Pretrain the model with the following command. Meanwhile there are pre-trained models, you can download them from here.

export XLA_PYTHON_CLIENT_PREALLOCATE=False
CUDA_VISIBLE_DEVICES=0 python launcher/examples/train_pretrain.py --variant 0 --seed 0

LoRA finetune

Train the skill policies with LoRA to achieve skill expansion. Meanwhile there are pre-trained models, you can download them from here.

CUDA_VISIBLE_DEVICES=0 python launcher/examples/train_lora_finetune.py --com_method 0 --model_cls 'LoRALearner' --variant 0 --seed 0

Context-aware Composition

Train the context-aware modular to adaptively leverage different skill knowledge to solve the tasks. You can download the pretrained model and datasets from here. Then, run the following command,

CUDA_VISIBLE_DEVICES=0 python launcher/examples/train_lora_finetune.py --com_method 0 --model_cls 'LoRASLearner' --variant 0 --seed 0

Citations

If you find our paper and code useful for your research, please cite:

@inproceedings{
liu2025psec,
title={Skill Expansion and Composition in Parameter Space},
author={Tenglong Liu, Jianxiong Li, Yinan Zheng, Haoyi Niu, Yixing Lan, Xin Xu, Xianyuan Zhan},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=GLWf2fq0bX}
}

Acknowledgements

Parts of this code are adapted from IDQL.