Training-free Task-oriented Grasp Generation

This project integrates Contact GraspNet and ManiSkill to enable a robotic system to generate task-oriented grasps without training. The system consists of two main components:

1. Contact GraspNet - Generates grasp candidates based on object geometry.
2. ManiSkill - Controls the robotic arm to execute grasping actions based on task requirements.

---

Getting Started

Clone the Repository

git clone https://github.com/your_username/Training-free-Task-oriented-Grasp-Generation.git
cd Training-free-Task-oriented-Grasp-Generation

---

Dependencies

Make sure the following dependencies are installed in their respective environments:

cgn_env (for grasp generation)

conda env create -f cgn_environment.yml

mani3_env (for robot control)

conda env create -f mani3_environment.yml

---

Setup Instructions

Step 1: Setup Conda Environments

Ensure you have Conda installed. Then, activate the required environments:

Grasp Generation (Contact GraspNet)

conda activate cgn_env

Robot Control (ManiSkill)

conda activate mani3

---

Step 2: Start the Contact GraspNet Server

In the first terminal, navigate to the Contact GraspNet directory and launch the grasping server:

cd ./third_party/contact_graspnet/contact_graspnet/
python socket_server.py

This starts a socket server that generates task-oriented grasping points.

---

Step 3: Run the Robot Controller

In another terminal, activate the ManiSkill environment and execute the robot control script:

conda activate mani_env2
python move_robot.py --task task_1 --model gemini --task_idx 1 --visual_method gripper_one

Command Arguments:

Argument	Description
--task task_1	Specifies the task (e.g., gripping toys).
--model gemini	Model used for grasp evaluation.
--task_idx 1	Index of the task instance.
--visual_method gripper_one	Defines the visualization method for grasp execution.