iCEM MPC Expert

Introduction A framework for robotic grasping using iterative Cross-Entropy Method (iCEM) with Model Predictive Control (MPC). Uses DexGraspNet hand positions for generate optimal approach trajectory.

Visualizations

Approach trajectory and grasp pose matching:

Features

• Implementation of iCEM algorithm for robotic control
• MPC-based grasp planning and execution
• Support for multiple object types (bottles, bowls, mugs, etc.)
• Integration with DexGraspNet dataset
• Physics-based simulation environment compatible with RL

Project Structure

.
├── app/                     # Example applications
├── icem_mpc/               # Core implementation
├── mjcf/                   # MuJoCo model files
├── final_positions/        # Pre-computed grasp positions
└── typings/                # Type definitions

Installation

1. Create environment using conda:

conda env create -f environment.yml

2. Install the package:

pip install -e .

Usage

Check the example scripts in app/ folder:

• example_icem_mpc.py - Basic iCEM-MPC implementation
• example_env.py - Environment setup example
• show_start_poses.py - Visualization of initial poses

Reward

The reward function in reach_pose_env.py encourages the robot hand to minimize the distance between key points on hand and the target grasp pose. A higher reward is given for precise alignment and successful approach, while penalties are applied for bumping an object.

Thanks

DexGraspNet -- for good postions

pytorch_icem -- for fast and simple implemnatation iCEM