Skip to content

we modified the ACT’s architecture to meet the unique kinematic and sensory requirements of the Franka robot.

Notifications You must be signed in to change notification settings

sainavaneet/ACTfranka

Repository files navigation

ACTFranka: Action chunk trandformer on franka robot

Welcome to the ACTFranka repository, designed for the Franka robot. This guide covers both simulated and real-world environment setups, and includes utilities for environment setup, training, and inference.

We have modified the original ACT code from this repo to complete this project, enhancing its capabilities to better suit our specific application needs.

website - https://sainavaneet.github.io/ACTfranka.github.io/

📋 Prerequisites

Ensure you have the following installed:

  • Ubuntu 20.04
  • ROS Noetic
  • libfranka package for Franka robot control

🚀 Installation

To get started with the ACT imitation learning framework, follow these steps:

git clone https://github.com/sainavaneet/ACTfranka.git
cd ACTfranka
pip install -r requirments.txt

🗂 Project Structure

  • controller/: Contains robot control code modules including movement and state management.
  • settings/: Houses configuration files for dataset paths and hyperparameters.
  • simulation/: Scripts for recording episodes and evaluating models are here.
  • train.py: The main script to train the policy using ACT imitation learning.
  • real_robot/: Specialized scripts for deploying the model on an actual Franka robot.

🏗 Step 1: Create the Environment

  1. Setup a simulated environment in Gazebo using the Franka robot and the libfranka package.
  2. Record episodes using the script located at simulation/record_episodes.py.
    • Make sure the dataset path is correctly set in settings/var.py.

Dataset Format

The dataset should be structured in HDF5 format as follows:

HDF5 file contents:
- action: <HDF5 dataset "action": shape (149, 8), type "<f8">
- observations:
  - images:
    - top: <HDF5 dataset "top": shape (149, 480, 640, 3), type "|u1">
  - qpos: <HDF5 dataset "qpos": shape (149, 8), type "<f8">

Replay Episodes

Utilize the Jupyter notebook dataset_prepare/replay.ipynb to replay recorded episodes by specifying the episode path.

🏋️ Step 2: Train the Model

  1. Configure the necessary hyperparameters in settings/var.py.
  2. Execute the train.py script with the prepared dataset to generate the policy.

🤖 Step 3: Model Inference

  1. Load and evaluate the trained policy using the script simulation/evaluate.py.
  2. This process simulates how the Franka robot will perform the learned tasks in a controlled environment.

🌍 Real Robot Deployment

To deploy on a real Franka robot, navigate to the real_robot directory. Scripts here are specifically adapted for real-world operations of the Franka robot.

🆘 Support

For any issues or further questions, please open an issue on the GitHub repository.

About

we modified the ACT’s architecture to meet the unique kinematic and sensory requirements of the Franka robot.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published