LaundryNauts/smolvla_augmented

LaundryNauts — LeHome Challenge 2026 Submission

Team: LaundryNauts | Registration ID: r20 | Institution: Northwestern University

Method

We fine-tune SmolVLA on 1,000 bimanual garment-folding episodes collected in the LeHome simulator. The model takes three RGB camera views (top, left wrist, right wrist) and 12-DOF joint state as input, and outputs 12-DOF dual-arm joint angle commands.

• Policy checkpoint + source: LaundryNauts/smolvla_augmented (HuggingFace)
• Training dataset: LaundryNauts/AugmentedDataset (HuggingFace)

---

Evaluation Setup

Requirements

• Linux x86_64, Python 3.11, uv package manager
• CUDA GPU
• Isaac Sim 5.1.0 (via the LeHome base repo)

Step 1 — Install the LeHome Base Environment

Follow the official LeHome setup instructions to install Isaac Lab and the lehome package.

Step 2 — Authenticate with HuggingFace

huggingface-cli login --token <hf_token_provided_in_submission>

Step 3 — Download Policy, Source Code, and Dataset Metadata

# Policy checkpoint
huggingface-cli download LaundryNauts/smolvla_augmented --local-dir outputs/smolvla_new

# Custom eval scripts (place into your lehome repo)
huggingface-cli download LaundryNauts/smolvla_augmented \
    scripts/eval_policy/lerobot_policy.py \
    scripts/eval_policy/base_policy.py \
    scripts/eval_policy/registry.py \
    --local-dir .

# Dataset metadata (only meta/ is required, not the full data)
huggingface-cli download LaundryNauts/AugmentedDataset \
    --repo-type dataset --local-dir Datasets/augmented_data

Step 4 — Download Sim Assets

huggingface-cli download lehome/asset_challenge --repo-type dataset --local-dir Assets

Step 5 — Run Evaluation

Single garment type:

export TMPDIR=/tmp/$(whoami) && mkdir -p $TMPDIR

LIVESTREAM=2 python -m scripts.eval \
    --policy_type lerobot \
    --policy_path outputs/smolvla_new \
    --garment_type top_long \
    --dataset_root Datasets/augmented_data \
    --num_episodes 2

All garment types:

for type in top_long top_short pant_long pant_short; do
    LIVESTREAM=2 python -m scripts.eval \
        --policy_type lerobot \
        --policy_path outputs/smolvla_new \
        --garment_type $type \
        --dataset_root Datasets/augmented_data \
        --num_episodes 2 \
        2>&1 | tee -a results_all.txt
done

---

Self-Reported Results

See results.txt for full per-garment breakdown.

Category	Success Rate
Top Long	16/24 = 66.7%
Top Short	6/24 = 25.0%
Pant Long	13/24 = 54.2%
Pant Short	18/24 = 75.0%
Overall	53/96 = 55.2%