tsingzeyong/lehome_4_30

LeHome Challenge Submission

This repository contains the Hugging Face submission package for the LeHome Challenge using the SmolVLA checkpoint trained on the merged four-garment dataset.

Contact

• Contact email: 820834523@qq.com

Included Contents

• Assets/ - bundled official garment assets required for evaluation
• checkpoints/pretrained_model/ - exported LeRobot checkpoint for evaluation
• scripts/ - evaluation entrypoint and policy loading code
• source/lehome/ - local LeHome package required by the evaluation code
• pyproject.toml - project dependencies
• configs/train_smolvla_four_types_3k.yaml - reference training config for this submission
• Datasets/example/four_types_merged/meta/ - dataset metadata required for LeRobot policy evaluation
• vlm/SmolVLM2-500M-Video-Instruct/ - bundled SmolVLM config/tokenizer/processor files required by the checkpoint

Policy Information

• Policy type: lerobot
• Policy backbone: smolvla
• Checkpoint path: checkpoints/pretrained_model
• Dataset metadata path: Datasets/example/four_types_merged/meta
• Recommended dataset root for evaluation: Datasets/example/four_types_merged
• Training run: 3k-step SmolVLA run on the merged four-garment dataset

Step-by-step setup

1. Clone or download this repository

git clone https://huggingface.co/tsingzeyong/lehome_4_30
cd lehome_4_30

2. Prepare the official evaluation environment

Use the official LeHome base environment. This submission was verified against the official stack:

• Python 3.11
• Isaac Sim 5.1
• LeRobot 0.4.3

At minimum:

python --version  # expected: Python 3.11.x
pip install -e source/lehome
pip install -e .

If the official evaluation workspace is already prepared, place this repository inside that workspace and run evaluation there.

This submission already includes the model checkpoint, the dataset meta files required by LeRobot, and the bundled SmolVLM tokenizer/config files required at load time.

3. Ensure asset files are reachable

This updated submission already bundles Assets/objects/Challenge_Garment.

You can either:

1. Run directly inside this submission repository.
2. Or copy this submission package into the official LeHome base workspace and run it there.

If the garment assets are stored elsewhere, set:

export LEHOME_GARMENT_CFG_BASE_PATH=/absolute/path/to/Assets/objects/Challenge_Garment

4. Verify the packaged files

Before evaluation, verify that these paths exist:

ls checkpoints/pretrained_model
ls vlm/SmolVLM2-500M-Video-Instruct
ls Datasets/example/four_types_merged/meta

Evaluation

5. Run evaluation

Example single-category evaluation command:

python -m scripts.eval \
    --policy_type lerobot \
    --policy_path checkpoints/pretrained_model \
    --garment_type top_long \
    --dataset_root Datasets/example/four_types_merged \
    --num_episodes 2 \
    --enable_cameras \
    --device cpu

The same checkpoint can be evaluated with top_long, top_short, pant_long, or pant_short.

Example loop over all four categories:

for garment in top_long top_short pant_long pant_short; do
    python -m scripts.eval \
        --policy_type lerobot \
        --policy_path checkpoints/pretrained_model \
        --dataset_root Datasets/example/four_types_merged \
        --garment_type "$garment" \
        --num_episodes 1 \
        --enable_cameras \
        --headless \
        --device cuda
done

For evaluation, only the metadata under Datasets/example/four_types_merged/meta/ is included in this repository. If your evaluator expects a different dataset location, keep the same meta/ contents available there.

Notes

• --policy_path must point to checkpoints/pretrained_model.
• Keep Datasets/example/four_types_merged/meta/ even if the raw training data is removed.
• The bundled vlm/SmolVLM2-500M-Video-Instruct/ directory is required because the checkpoint references it during loading.
• Relative VLM and asset paths are resolved automatically from the submission root; no machine-local /root/... paths are required.
• No private credentials are required for evaluation.

Upload to Hugging Face

From inside this submission folder:

hf auth login
hf upload tsingzeyong/lehome_4_30 . --repo-type model