Probe_Exploration

This repository provides a single-GPIS Isaac Lab testbed for our paper
“Autonomous Exploration for Shape Reconstruction and Measurement via Informative Contact-Guided Planning (AESRM)”.

🔗 Project Page: https://aesrm.github.io

It implements a simplified version of the AMMH framework in simulation, focusing on tactile exploration and surface reconstruction with a simple contact points, and is mainly intended for testing and debugging the GPIS-based exploration pipeline.

Project Structure

Hand_Exploration/
├── Results/
├── Env/
│   ├── allegro_hand.py               # Allegro Hand model definitions
│   ├── Explloration_env.py            # Tactile exploration environment setup (Stage 1)
│   ├── Exploration_env_stage2.py      # Advanced exploration environment (Stage 2)
│   ├── ....
├── utils/
│   ├── GPIS.py                        # Basic GPIS modeling
│   ├── HE_GPIS.py                     # Hand Exploration-specific GPIS extensions
│   ├── Pid_Controller.py              # PID control for hand pose
│   ├── ....
├── README.md                          

Installation

1. Clone the repository:

   git clone https://github.com/FerryRain/Hand_Exlporation.git
   cd Hand_Exploration

2. Create the Conda environment:

   Create a new environment using the provided environment.yaml:

   conda env create -f environment.yaml
   conda activate hand_exploration

3. Install Isaac Sim and Isaac Lab dependencies manually:

   • Follow the official installation guides:

     • Isaac Sim Installation Guide
     • Isaac Lab Setup Instructions

   • Typically, you need to:

     • Download and install Isaac Sim.

     • Set up environment variables (e.g., source setup_env.sh inside Isaac Sim folder).

     • Install Isaac Lab Python packages inside your environment:

       cd /path/to/isaaclab
       pip install -e .

   ⚠️ Note:
   Isaac Sim and Isaac Lab require specific versions of CUDA, GPU drivers, and system dependencies.
   Make sure your machine meets their requirements.

Main Components

• Env/

  • allegro_hand.py: Robot hand model and kinematics definition.
  • Explloration_env.py: Basic tactile exploration simulation environment.
  • Exploration_env_stage2.py: Extended environment for Stage 2 exploration tasks, incorporating surface reconstruction and path planning.

• utils/

  • GPIS.py: Standard Gaussian Process Implicit Surface (GPIS) modeling.
  • HE_GPIS.py: Hand Exploration-enhanced GPIS with uncertainty-guided exploration strategies.
  • Pid_Controller.py: Proportional-Integral-Derivative (PID) controller for controlling hand position and orientation.

• TEST/

  • Reserved for test scripts and experimental trials.

Getting Started

1. Run the simple exploration demo:

   conda activate hand_exploration

   cd Env
   python Exploration_env_stage{x}.py

---

Experiment Stages Overview

The project progresses through nine stages of tactile exploration and surface reconstruction:

1. Stage 1: Teleportation + Contact

   • The hand is directly teleported to object surface to collect initial contact points.

2. Stage 2: move_to + Uncertainty

   • The hand moves toward regions of higher surface uncertainty using basic motion primitives.

3. Stage 3: Surface Constraint + Smoothness

   • Hand motion is constrained on the object surface.
   • Guided by both uncertainty gradient and contact smoothness (possibly using δ constraints).

4. Stage 4: Global Point Sampling as AS Input

   • Actively samples next exploration targets from global surface uncertainty map.

5. Stage 5: Position Control Version

   • Replaces high-level move_to with low-level position control for trajectory execution.

6. Stage 6: Continuous PID Control

   • Employs PID-based continuous control to smoothly follow desired paths along surface.

7. Stage 7: New GPIS Initialization Function

   • Improves initialization of the GPIS model with different spatial priors (e.g. spherical or conic).

8. Stage 8: Hybrid Force-Position Control

   • Combines force feedback with position control to ensure compliant, stable contact.

---

Results Organization

Experimental data for each stage is stored in the Data/ directory with folders named accordingly. Some examples:

Results/
├── stage6_1/                     # Continuous control (PID) for sphere (radius=0.5) (falied)
├── stage6_2/                     # Continuous control (PID) for sphere (radius=0.15)
├── stage6_3(CONE)/               # Continuous control (PID) for cone (radius=0.15, height=0.3)
├── stage7_1_1(sphere)/           # Stage 7 for GPIS which initialized by double bb for sphere (radius=0.15)
├── stage7_1_2/                   # Stage 7 for GPIS which initialized by double bb for sphere (radius=0.5) (success)
├── stage7_2_1/                   # Stage 7 for GPIS which initialized by double bb for cone (radius=0.15, height=0.3)
├── stage7_2_2/                   # Stage 7 for GPIS which initialized by double bb for cone (radius=0.15, height=0.3) with another type of the gaussian data preprocess 
├── stage8_1/                     # Stage 8 experiment with hybrid control for sphere
├── stage8_2_1/                   # Stage 8 experiment with hybrid control for cone

To visualize saved 3D reconstruction results:

cd Results
python Vis_points.py

---