ReSemAct: Advancing Fine-Grained Robotic Manipulation via Semantic Structuring and Affordance Refinement

[Project Page]

---

🛠️ Environment

• OS: Ubuntu 20.04
• CUDA: 12.2
• NVIDIA Driver: 535.161.07
• Conda Environments:
  • ReSemAct (Client)
  • m3p2i-aip (Server)

---

🔧 Client Setup (ReSemAct)

1. Create Conda Environment

conda create -n ReSemAct python=3.10 
conda activate ReSemAct

2. Install OmniGibson

Install from source (editable mode), version v1.1.0:

git clone -b v1.1.0 https://github.com/StanfordVL/BEHAVIOR-1K.git
cd BEHAVIOR-1K/

Install OmniGibson:

pip install -e .
python -m omnigibson.install

3. Test Installation

cd <BEHAVIOR-1K_folder>
python -m omnigibson.examples.robots.robot_control_example --quickstart

4. Clone the ReSemAct Repository

git clone https://github.com/scy-v/ReSemAct.git

5. Install FastSAM and Download Weights

cd ReSemAct
git clone https://github.com/CASIA-IVA-Lab/FastSAM.git

6. Install Additional Dependencies

pip install -r requirements_client.txt

---

🖥️ Server Setup (m3p2i-aip)

1. Clone the m3p2i-aip Repository and Create Environment

cd <ReSemAct_folder>
git clone https://github.com/tud-amr/m3p2i-aip.git
conda create -n m3p2i-aip python=3.8
conda activate m3p2i-aip

2. Install Isaac Gym

Follow the prerequisites guide, download Isaac Gym from NVIDIA:https://developer.nvidia.com/isaac-gym

Move and install Isaac Gym:

mv <Downloaded_Folder>/IsaacGym_Preview_4_Package <ReSemAct_folder>/m3p2i-aip/thirdparty/
cd <ReSemAct_folder>/m3p2i-aip/thirdparty/IsaacGym_Preview_4_Package/isaacgym/python
pip install -e .

3. Install m3p2i-aip

cd <ReSemAct_folder>/m3p2i-aip
pip install -e .

4. Install Additional Dependencies

cd <ReSemAct_folder>
pip install -r requirements_server.txt

---

🚀 Running the Demo

You need two terminals: one for the server and one for the client.

1. Start the Server

cd <ReSemAct_folder>
python mppi_server.py

2. Run the Client

Before running, set your OpenAI API key by either:

export OPENAI_API_KEY="your_api_key_here"

or add it to configs/omnigibson_config/config.yaml. Then start running with:

cd <ReSemAct_folder>
python run.py [--load_cache] [--visualize]

• --load_cache: Load pre-generated GPT-4o cache
• --visualize: Enable visual debugging

---

⚡Known Issues

1. When the Franka robot grasps the object, an error may sometimes occur at the moment of contact, causing the sticky grasp mode to fail and the object to slip out of the gripper.

2. Due to simulation limits (execute_action), the robot’s optimization and execution run at a low frequency (~6 Hz). If the robot struggles to reach the target pose, please adjust threshold or deploy in a real environment.

---

🔑 Acknowledgments

• Simulation Environments
  The simulation environments are based on OmniGibson and Isaac Gym.

• Language Model Integration
  The extension of Language Model Programs (LMPs) is built upon Voxposer and Code as Policies.

• Motion Planning
  The Model Predictive Path Integral (MPPI) algorithm implemented on Isaac Gym is adopted from m3p2i-aip.

• Code Snippets Reference
  Part of the environment code is adapted from the ReKep project.