ContactExplorer

Highlights

ContactExplorer is a exploration method for dexterous manipulation. It defines contact as the intersection between object surface points and hand keypoints, and maintains a hash-conditioned counter of which fingers touch which object regions.

• Coverage reward (count-based) rewards novel contact patterns.
• Reaching reward (energy-based) steers the hand toward under-explored regions.
• Results: faster training and higher success on singulation, retrieval, in-hand reorientation, and bimanual tasks — with sim-to-real transfer. See the paper.

📚 Table of Contents

1. Overview
2. Installation
   • IsaacGym Conda Env
   • Install IsaacGym
   • Install Other Dependencies
3. Training and Evaluation
   • Training Scripts
   • Evaluation
   • Observation and Action Spaces
   • Reward Settings (reward_type)
4. Repository Structure
5. CCGE Reward Architecture
6. Citation
7. License

Installation

IsaacGym Conda Env

conda create -n ccge python=3.8   # mamba also works
conda activate ccge

Install IsaacGym

Download IsaacGym and extract:

wget https://developer.nvidia.com/isaac-gym-preview-4
tar -xvzf isaac-gym-preview-4

Install IsaacGym Python API:

pip install -e isaacgym/python

Test installation:

python 1080_balls_of_solitude.py  # or
python joint_monkey.py

For libpython error:

• Check conda path:

  conda info -e

• Set LD_LIBRARY_PATH:

  export LD_LIBRARY_PATH=</path/to/conda/envs/your_env/lib>:$LD_LIBRARY_PATH

Install Other Dependencies

Install IsaacGymEnvs and following dependencies:

pip install --no-build-isolation -r requirements.txt

Training and Evaluation

Training Scripts

Two types of dexterous hands are provided (LEAP and Allegro). You may choose one of them to train.

Task	Training Script
Singulation	train_<hand_type>_singulation.sh
Table Top	train_<hand_type>_table_top.sh
Inhand	train_<hand_type>_inhand.sh
Retri	train_<hand_type>_cube_in_box.sh
Bimanual	train_bimanual.sh

Evaluation

Set mode=eval and point to a trained run directory. Start from the corresponding train_*.sh and append the eval flags:

python src/train.py \
    mode=eval \
    task=<TaskName> \
    train=<TrainCfgName> \
    ... \
    --model_dir=logs/PPO/<run_dir> \
    --resume_iter=<checkpoint_iter> \
    --eval_times=5 \
    --vis_env_num=0

Notes:

• --model_dir is required for evaluation and should contain model_*.pt.
• --resume_iter is optional (defaults to the latest checkpoint).

Observation and Action Spaces

Observation and action spaces are set via Hydra overrides in launch scripts:

obs_space="['allegro_hand_dof_position']"
action_space="['wrist_translation','wrist_rotation','hand_rotation']"

Available keys are task-specific — see the corresponding file in src/tasks/.

Reward Settings (reward_type)

Training scripts pass a reward_type string to src/train.py, e.g.:

• reward_type="target+bonus+success+reach+energy_reach+contact_coverage"

The CCGE exploration signal consists of energy_reach and contact_coverage. To ablate exploration, remove them:

• reward_type="target+bonus+success+reach"

Repository Structure

• src/ — core library code
  • tasks/ — Isaac Gym task environments, reward logic, and curiosity modules
  • algorithms/ — PPO and intrinsic-reward components
  • utils/ — config loading, logging, helpers
  • Entry points: train.py
• cfg/ — Hydra configs
  • task/ — task/environment configs
  • train/ — training configs

CCGE Reward Architecture

graph TD
    subgraph Inputs
        KP["Hand Keypoints<br/><i>(L points from URDF)</i>"]
        PC["Canonical Object<br/>Point Cloud + Normals<br/><i>(M points)</i>"]
        SFB["State Feature Bank<br/><i>LearnedHashStateBank /<br/>PushBox2DStateBank</i>"]
    end

    PC -->|K-means + FPS| CL["Surface Clusters<br/><i>(K clusters)</i>"]
    PC --> CRM
    CL --> CRM
    KP --> CRM
    SFB -->|state ID| CRM

    subgraph CRM ["CuriosityRewardManager"]
        POT["Energy-based Reaching Reward Φ<br/><i>novelty-weighted kernel</i>"]
        CB["Contact Coverage Reward<br/><i>cluster novelty</i>"]
        RM["Running-Max Tracker<br/><i>per state × keypoint</i>"]
    end

    CRM --> REW["<b>CCGE Reward = Energy-based Reaching Reward + Contact Coverage Reward</b>"]

Loading

Required Data

Each task must supply these tensors per step (N = num envs, L = keypoints, M = object points):

Tensor	Shape	How to get it
keypoint_positions_with_offset	(N, L, 3)	Index rigid-body states by keypoint link indices, apply local offsets via quat_apply
keypoint_contact_mask	(N, L) bool	(dist_to_surface < threshold) & (contact_force > threshold)
object_root_positions	(N, 3)	From root_states
object_root_orientations	(N, 4)	From root_states (xyzw quaternion)
Canonical point cloud	(M, 3)	Loaded from dataset (object frame)
Canonical normals	(M, 3)	Loaded from dataset

For the full step-by-step integration guide (keypoint setup, contact mask, CuriosityRewardManager init, reward computation, reset handling, and config), see src/tasks/README.md.

Acknowledgments

This repository builds upon or incorporates code from the following open-source projects:

• UniDexFPM for hand-arm task environments.
• IsaacGymEnvs for base task environments and Isaac Gym utilities.
• WoCoCo for intrinsic baseline implementation references.
• ARCTIC and ContactDB for simulation assets.

Please refer to the respective repositories and their licenses for more details.

Citation

If you find our work useful, please consider citing us!

@article{liu2026contactcoverageguidedexplorationgeneralpurpose,
      title={Contact Coverage-Guided Exploration for General-Purpose Dexterous Manipulation}, 
      author={Zixuan Liu and Ruoyi Qiao and Chenrui Tie and Xuanwei Liu and Yunfan Lou and Chongkai Gao and Zhixuan Xu and Lin Shao},
      year={2026},
      journal={arXiv preprint arXiv:2603.10971},
}

License

This project is licensed under the MIT License - see the LICENSE file for details.