zhenyuzhao/H_RDT

H-RDT: Human Manipulation Enhanced Bimanual Robotic Manipulation

📝Paper | 🌍Project Page | 🤗Model | 💬WeChat Contact

📰 News

• [2025.8.12] Updated RoboTwin2 inference code

H-RDT (Human to Robotics Diffusion Transformer) is a novel approach that leverages large-scale egocentric human manipulation data to enhance robot manipulation capabilities. Our key insight is that large-scale egocentric human manipulation videos with paired 3D hand pose annotations provide rich behavioral priors that capture natural manipulation strategies and can benefit robotic policy learning.

🚀 Installation

1. Create conda environment:

   conda create -n hrdt python=3.10
   conda activate hrdt
   

2. Install dependencies:

   pip install -r requirements.txt
   

3. Download pre-trained models:

   export HF_ENDPOINT=https://hf-mirror.com
   huggingface-cli download --resume-download embodiedfoundation/H-RDT --local-dir ./
   

🔧 Usage

Stage 1: Human Data Pre-training (EgoDx)

Data Preprocessing

Before training, preprocess the EgoDx dataset:

1. Configure paths:

   # Edit datasets/pretrain/setup_pretrain.sh with your paths
   nano datasets/pretrain/setup_pretrain.sh
   
   # Set your EgoDx dataset and T5 model paths:
   export EGODEX_DATA_ROOT="/path/to/your/egodx/dataset"
   export T5_MODEL_PATH="/path/to/your/t5-v1_1-xxl"
   

2. Setup environment:

   source datasets/pretrain/setup_pretrain.sh
   

3. Run data processing pipeline:

   # Automatically runs: precompute_48d_actions.py → calc_stat.py → encode_lang_batch.py
   ./datasets/pretrain/run_pretrain_pipeline.sh
   

Start Pre-training

After data preprocessing is complete:

1. EgoDx Pretrain (fresh start):

1. Configure dataset:

   # Edit datasets/dataset.py line ~45
   self.dataset_name = "egodx"
   

2. Run training:

   bash pretrain.sh
   

2. Pretrain Resume: Edit pretrain.sh, add this line:

--resume_from_checkpoint="checkpoint-450000" \

Stage 2: Cross-Embodiment Fine-tuning

Data Preprocessing (for RobotWin2)

Pre-computed language embeddings are already provided - no preprocessing needed!

1. Setup environment:

   # Edit datasets/robotwin2/setup_robotwin2.sh if needed (only for regenerating files)
   source datasets/robotwin2/setup_robotwin2.sh
   

2. Data processing pipeline (Not Required):

   # Not needed - lang_embeddings/ already provided in repository
   # Only run if you want to regenerate files:
   # ./datasets/robotwin2/run_robotwin2_pipeline.sh
   

Robot Fine-tuning (load human pre-trained backbone):

1. Configure dataset:

   # Edit datasets/dataset.py line ~45
   self.dataset_name = "robotwin_agilex"  # or your robot name
   
   # Add your dataset initialization if not exists:
   elif self.dataset_name == "your_robot":
       self.hdf5_dataset = YourRobotDataset(config=config)
   

2. Run training:

   bash finetune.sh  # Already configured with pretrained_backbone_path
   

Finetune Resume:

Edit your current finetune script, make these changes:

# Change this line:
--mode="finetune" \
# To:
--mode="pretrain" \

# And add:
--resume_from_checkpoint="checkpoint-5000" \

🎯 Training Modes

Training Scenario	Base Script	Required Shell Script Modifications	Mode & Key Parameters
Human Pretrain (Fresh)	pretrain.sh	--mode="pretrain"	Start pretraining on EgoDx human data
Human Pretrain Resume	pretrain.sh	Add: --resume_from_checkpoint="checkpoint-450000" \	--mode="pretrain"
Robot Fine-tuning	finetune.sh	Change: --mode="finetune" \ Add: --pretrained_backbone_path="./checkpoints/pretrain-0618/checkpoint-500000/pytorch_model.bin" \ Change: --config_path="configs/hrdt_finetune.yaml" \	Load human pre-trained backbone, fresh action layers
Robot Finetune Resume	Your finetune script	Change: --mode="finetune" → --mode="pretrain" Add: --resume_from_checkpoint="checkpoint-5000" \	Continue robot fine-tuning

Dataset Configuration

Before training, you can select datasets via CLI flags or configure in datasets/dataset.py:

For Human Pre-training (EgoDx):

--dataset_name=egodex

For Robot Fine-tuning:

--dataset_name=robotwin_agilex

Adding New Robot Datasets:

1. Create your dataset folder: datasets/your_robot/
2. Implement your dataset class (see datasets/robotwin2/ as example)
3. Create data processing scripts (see datasets/pretrain/ or datasets/robotwin2/ as examples)
4. Import in datasets/dataset.py
5. Add initialization logic in VLAConsumerDataset.__init__

LeRobot (HE -> LeRobot) Fine-tuning:

--dataset_name=lerobot --dataset_root=/hfm/data/pick_n_squat

See datasets/lerobot/README.md for action stats and language embeddings setup.

Key Configuration Files

• configs/hrdt_pretrain.yaml: Human pre-training configuration
• configs/hrdt_finetune.yaml: Robot fine-tuning configuration
• datasets/dataset.py: Dataset selection and initialization
• Modify state_dim, action_dim, output_size for your robot

📞 Contact Us

WeChat Discussion Group

Join our WeChat group to discuss H-RDT related technical issues:

WeChat Group QR Code

Personal WeChat

For other questions or collaboration opportunities, please add personal WeChat:

Personal WeChat QR Code

---

Note: If the QR code expires, please contact us through project Issues for the latest contact information.