PSEC
[ICLR 2025] The offical implementation of "PSEC: Skill Expansion and Composition in Parameter Space", a new framework designed to facilitate efficient and flexible skill expansion and composition, iteratively evolve the agents' capabilities and efficiently address new challenges
Install / Use
/learn @ltlhuuu/PSECREADME
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.
Related Skills
clearshot
Structured screenshot analysis for UI implementation and critique. Analyzes every UI screenshot with a 5×5 spatial grid, full element inventory, and design system extraction — facts and taste together, every time. Escalates to full implementation blueprint when building. Trigger on any digital interface image file (png, jpg, gif, webp — websites, apps, dashboards, mockups, wireframes) or commands like 'analyse this screenshot,' 'rebuild this,' 'match this design,' 'clone this.' Skip for non-UI images (photos, memes, charts) unless the user explicitly wants to build a UI from them. Does NOT trigger on HTML source code, CSS, SVGs, or any code pasted as text.
openpencil
2.0kThe world's first open-source AI-native vector design tool and the first to feature concurrent Agent Teams. Design-as-Code. Turn prompts into UI directly on the live canvas. A modern alternative to Pencil.
ui-ux-designer
Use this agent when you need to design, implement, or improve user interface components and user experience flows. Examples include: creating new pages or components, improving existing UI layouts, implementing responsive designs, optimizing user interactions, building forms or dashboards, analyzing existing UI through browser snapshots, or when you need to ensure UI components follow design system standards and shadcn/ui best practices.\n\n<example>\nContext: User needs to create a new dashboard page for team management.\nuser: "I need to create a team management dashboard where users can view team members, invite new members, and manage roles"\nassistant: "I'll use the ui-ux-designer agent to design and implement this dashboard with proper UX considerations, using shadcn/ui components and our design system tokens."\n</example>\n\n<example>\nContext: User wants to improve the user experience of an existing form.\nuser: "The signup form feels clunky and users are dropping off. Can you improve it?"\nassistant: "Let me use the ui-ux-designer agent to analyze the current form UX and implement improvements using our design system and shadcn/ui components."\n</example>\n\n<example>\nContext: User wants to evaluate and improve existing UI.\nuser: "Can you take a look at our pricing page and see how we can make it more appealing and user-friendly?"\nassistant: "I'll use the ui-ux-designer agent to take a snapshot of the current pricing page, analyze the UX against Notion-inspired design principles, and implement improvements using our design tokens."\n</example>
HappyColorBlend
HappyColorBlendVibe Project Guidelines Project Overview HappyColorBlendVibe is a Figma plugin for color palette generation with advanced tint/shade blending capabilities. It allows designers to
