Skip to content

Latest commit

 

History

History
184 lines (133 loc) · 7.73 KB

README.md

File metadata and controls

184 lines (133 loc) · 7.73 KB

LeHome Challenge 2026

Challenge on Garment Manipulation Skill Learning in Household Scenarios

Competition Website

Python Isaac Lab LeRobot License ICRA

📑 Table of Contents

🚀 Quick Start

⚠️ IMPORTANT: For Ubuntu version and GPU-related settings, please refer to the IsaacSim 5.1.0 Documentation. And the simulation currently only supports CPU devices.

1. Installation

We offer two installation methods: UV and Docker for submission and local evaluation.

Use UV

The simulation environment is based on the IssacLab and LeRobot repositories; please refer to UV installation guide.

Use Docker

The simulation environment is based on the IssacLab and LeRobot repositories; please refer to Docker installation guide.

2. Assets & Data Preparation

Download Simulation Assets

Download the required simulation assets (scenes, objects, robots) from HuggingFace:

# This creates the Assets/ directory with all required simulation resources
hf download lehome/asset_challenge --repo-type dataset --local-dir Assets

Download Example Dataset

We provide demonstrations for four types of garments. Download from HuggingFace:

hf download lehome/dataset_challenge_merged --repo-type dataset --local-dir Datasets/example

If you need depth information or individual data for each garment. Download from HuggingFace:

hf download lehome/dataset_challenge --repo-type dataset --local-dir Datasets/example

Collect Your Own Data

For detailed instructions on teleoperation data collection and dataset processing, please refer to our Dataset Collection and Processing Guide ( using SO101 Leader is strongly recommended).

3. Train

We provide several training examples; the models and training framework are from LeRobot.

Quick Start

Train using one of the pre-configured training files:

lerobot-train --config_path=configs/train_<policy>.yaml

Available config files:

  • configs/train_act.yaml - ACT
  • configs/train_dp.yaml - DP
  • configs/train_smolvla.yaml - SmolVLA

Key configuration options:

  • Dataset path: Update dataset.root to point to your dataset
  • Input/Output features: Specify which observations and actions to use
  • Training parameters: Adjust batch_size, steps, save_freq, etc.
  • Output directory: Modify output_dir to save models elsewhere

📖 For detailed training instructions, feature selection guide, and configuration options, see our Training Guide.

4. Eval

Evaluate your trained policy on the challenge garments. The framework supports LeRobot policies and custom implementations.

Examples:

# Evaluate using LeRobot policy 
# Note: --policy_path and --dataset_root are required parameters for LeRobot policies, ready to run once the dataset and model checkpoints are prepared.
python -m scripts.eval \
    --policy_type lerobot \
    --policy_path outputs/train/act_top_long/checkpoints/last/pretrained_model \
    --garment_type "top_long" \
    --dataset_root Datasets/example/top_long_merged \
    --num_episodes 2 \
    --enable_cameras \
    --device cpu    

# Evaluate custom policy
# Note: Participants can define their own model loading logic within the policy class. Provides flexibility for participants to implement specialized loading and inference logic.
python -m scripts.eval \
    --policy_type custom \
    --garment_type "top_long" \
    --num_episodes 5 \
    --enable_cameras \
    --device cpu

Common Options

Parameter Description Default Required For
--policy_type Policy type: lerobot, custom lerobot All
--policy_path Path to model checkpoint - All (passed as model_path for custom)
--dataset_root Dataset path (for metadata) - LeRobot only
--garment_type Type of garments: top_long, top_short, pant_long, pant_short, custom top_long All
--num_episodes Episodes per garment 5 All
--max_steps Max steps per episode 600 All
--save_video Save evaluation videos All
--video_dir Directory to save evaluation videos outputs/eval_videos --save_video
--enable_cameras Enable camera rendering All
--device Device for inference: only cpu 'cpu' All
--headless Used for evaluation without GUI disabled All

Parameter Descriptions:

  • Required for LeRobot Policy: --policy_path (model path) and --dataset_root (dataset path, used for loading metadata).
  • Custom Policy: --policy_path is passed to the policy constructor as model_path. Participants can define their own model loading logic (refer to scripts/eval_policy/example_participant_policy.py).

Garment Test Configuration

Evaluation is performed on the Release set of garments. Under the directory Assets/objects/Challenge_Garment/Release, each garment category folder contains a corresponding text file listing the garment names (e.g., Top_Long/Top_Long.txt).

  • Evaluate a Category: Set --garment_type to top_long, top_short, pant_long, or pant_short to evaluate all garments within that category.
  • Evaluate Specific Garments: Edit Assets/objects/Challenge_Garment/Release/Release_test_list.txt to include only the garments you want to test, then run with --garment_type custom.

📖 For detailed policy evaluation guide, see eval_guide.

📮 Submission

Once you are satisfied with your model's performance, follow these steps to submit your results to the competition leaderboard:

Submission instructions will be available on the competition website.

🧩 Acknowledgments

This project stands on the shoulders of giants. We utilize and build upon the following excellent open-source projects:

  • Isaac Sim - For photorealistic physics simulation
  • Isaac Lab - For modular robot learning environments
  • LeRobot - For state-of-the-art Imitation Learning algorithms
  • Marble - For diverse simulation scene generation