Geonhyup Lee, Youngjin Lee, Kangmin Kim, Seongju Lee, Sangjun Noh, Seunghyeok Back, Kyoobin Lee

---

This is an official implementation for "ManipForce: Force-Guided Policy Learning with Frequency-Aware Representation for Contact-Rich Manipulation", 2026 IEEE International Conference on Robotics and Automation (ICRA 2026).

🛠️ Setup

# 1. Install mamba (if not already installed)
conda install -c conda-forge mamba -n base -y

# 2. Create conda environment
mamba env create -f environment.yml

# 3. Activate environment
conda activate manipforce

# 4. Download pre-trained models
python checkpoints/prepare_dinov2.py --split-qkv

📡 Data Collection & Processing

# Step 1: Capture multimodal data
python scripts/collection/capture_multimodal_data.py --data_path data/<your_task> --add_cam

# Step 2: Synchronize multi-camera images
python scripts/collection/align_multimodal_data.py --data_path data/<your_task>

# Step 3: Estimate wrist marker pose
python scripts/processing/get_wrist_pose.py --data_path data/<your_task> --visualize

# Step 4: Refine pose with filtering and interpolation
python scripts/processing/pose_refinement.py --data_path data/<your_task>

# Step 5: Convert processed data to Zarr format
python scripts/processing/change_to_zarr.py --data_path /home/geonhyup/Workspace/ManipForce/data/11 --output_path /home/geonhyup/Workspace/ManipForce/data/11.zarr

🏋️ Training

Our method supports different observation down-sampling steps.

# Provide a predefined key or a direct path to a .zarr dataset
python scripts/launch.py --gpu 0 --config manipforce_ods3_256x256 --dataset data/your_task.zarr

🤖 Evaluation

python scripts/eval/eval_robot.py --config_path "eval_config/gear_insertion.yaml" 

📝 Arguments

Argument	Description	Default
--gpu	GPU ID to use for training.	0
--config	Hydra configuration file name (with or without .yaml).	manipforce_ods3_256x256
--dataset	Predefined key (e.g., gear, battery) or a direct path to a .zarr file.	Required