HIL-SERL: Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning

Webpage: https://hil-serl.github.io/

HIL-SERL provides a set of libraries, env wrappers, and examples to train RL policies using a combination of demonstrations and human corrections to perform robotic manipulation tasks with near-perfect success rates. The following sections describe how to use HIL-SERL. We will illustrate the usage with examples.

🎬: HIL-SERL video

Table of Contents

• HIL-SERL: Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning
  • Installation
  • Overview and Code Structure
  • Run with Franka Arm
  • Citation

Installation

1. Setup Conda Environment: create an environment with

   conda create -n hilserl python=3.10

2. Install Jax as follows:

   • For CPU (not recommended):

     pip install --upgrade "jax[cpu]"

   • For GPU:

     pip install --upgrade "jax[cuda12_pip]==0.4.35" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

   • For TPU

     pip install --upgrade "jax[tpu]" -f https://storage.googleapis.com/jax-releases/libtpu_releases.html

   • See the Jax Github page for more details on installing Jax.

3. Install the serl_launcher

   cd serl_launcher
   pip install -e .
   pip install -r requirements.txt

4. Install for serl_robot_infra Follow the README in serl_robot_infra for installation and basic robot operation instructions. This contains the instruction for installing the impendence-based serl_franka_controllers. After the installation, you should be able to run the robot server, interact with the gym franka_env (hardware).

Overview and Code Structure

HIL-SERL provides a set of common libraries for users to train RL policies for robotic manipulation tasks. The main structure of running the RL experiments involves having an actor node and a learner node, both of which interact with the robot gym environment. Both nodes run asynchronously, with data being sent from the actor to the learner node via the network using agentlace. The learner will periodically synchronize the policy with the actor. This design provides flexibility for parallel training and inference.

Table for code structure

Code Directory	Description
examples	Scripts for policy training, demonstration data collection, reward classifier training
serl_launcher	Main code for HIL-SERL
serl_launcher.agents	Agent Policies (e.g. SAC, BC)
serl_launcher.wrappers	Gym env wrappers
serl_launcher.data	Replay buffer and data store
serl_launcher.vision	Vision related models and utils
serl_robot_infra	Robot infra for running with real robots
serl_robot_infra.robot_servers	Flask server for sending commands to robot via ROS
serl_robot_infra.franka_env	Gym env for Franka robot

Run with Franka Arm

We provide a step-by-step guide to run RL policies with HIL-SERL on a Franka robot.

Check out the Run with Franka Arm

• RAM Insertion
• USB Pickup and Insertion
• Object Handover
• Egg Flip

Citation

If you use this code for your research, please cite our paper:

@misc{luo2024hilserl,
      title={Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning},
      author={Jianlan Luo and Charles Xu and Jeffrey Wu and Sergey Levine},
      year={2024},
      eprint={2410.21845},
      archivePrefix={arXiv},
      primaryClass={cs.RO}
}