README.md

metadata

license: apache-2.0
tags:
  - lerobot
  - lehome
  - act
  - robotics

ACT Policy For LeHome Challenge

This repository contains a LeRobot-format ACT checkpoint exported from:

• Experiment: outputs/train/act_top_longee_depth_bs32
• Checkpoint: outputs/train/act_top_longee_depth_bs32/checkpoints/0000125000
• Team: HandX
• Registration ID: r26
• Contact: xtli312@163.com

Expected Layout

• pretrained_model/
• train_act_1.yaml
• LEHOME_REPO_README.md

Evaluation

Use this checkpoint with the LeHome evaluation script:

python -m scripts.eval \
    --policy_type lerobot \
    --policy_path "<downloaded_bundle>/pretrained_model" \
    --dataset_root "Datasets/example/top_long_merged" \
    --garment_type "custom" \
    --num_episodes 5 \
    --task_description "fold the garment on the table" \
    --enable_cameras \
    --device cpu

Required Dependencies

• Python 3.11
• Isaac Sim 5.1.0
• Isaac Lab 2.3.1
• lerobot
• LeHome repository source code

Notes

• The challenge organizers also need the official assets and dataset metadata.
• If you upload this bundle to Hugging Face, point the Google Form Policy Submission field to that repository URL.
• Local self-reported evaluation on Release_test_list.txt with 5 episodes per garment:
  • Total Episodes: 15
  • Average Return: 183.60 +- 67.87
  • Success Rate: 20.00%
  • Top_Long_Seen_6: 20.00% success, 224.89 avg return
  • Top_Long_Seen_7: 0.00% success, 205.17 avg return
  • Top_Long_Unseen_1: 40.00% success, 120.74 avg return
• Generated on 2026-04-02T12:02:18.