UCAS-LeHome/policies

🤖 1st Lehome Challenge Submission

Team Name: Affiliation: UCAS Track: Simulation Challenge

---

1. 📑 Method Overview

Our solution implements a Dynamic Mixture-of-Experts (MoE) framework to address the multi-category garment manipulation task in the challenge.

• Vision-based Router: We utilize a fine-tuned ResNet18 classifier to determine the specific garment category (pant_long, pant_short, top_long, top_short) from the initial RGB observation frame.
• Specialized Experts: Each expert is trained and optimized exclusively on one garment category to ensure high-precision manipulation and robust handling of specific cloth geometries.
• Dynamic Policy Switching: At the start of each evaluation episode, the system identifies the target type and dynamically routes the control flow to the corresponding expert, ensuring the most suitable policy is active for the task at hand.

2. 🏠 Repository Structure

.
├── screenshots/                                        # Evaluation result screenshots
├── scripts/eval_policy/example_participant_policy.py  # Core MoE Logic & Dynamic Routing
├── outputs/classifier/garment_classifier_resnet18.pth # Fine-tuned ResNet18 Weights
├── outputs/train/                                     # Specialized Expert Weights
└── Datasets/example/                                  # Metadata for Normalization

3. 🧮 Evaluation Instructions

To evaluate our policy, please ensure the official Isaac Sim and LeRobot environment is properly configured. Our policy is registered under the type custom.

3.1 Retrieve Policy and Weights

Please clone this Hugging Face repository into your evaluation workspace:

git lfs install
git clone https://huggingface.co/UCAS-LeHome/policies

Please ensure that all files referenced under policies/ also exist at the corresponding paths under the original project root, or are mapped there via symbolic links.

3.2 Recommended Execution (By Category)

python -m scripts.eval --policy_type custom --garment_type custom --enable_cameras --device cpu

4. 📧 Contact Information

• Primary Contact: Zhanbo Wang (beachbum9527@gmail.com)