LaNE

Overview

Implementation of CoRL 2024 paper: LaNE: Accelerating Visual Sparse-Reward Learning with Latent Nearest-Demonstration-Guided Explorations

LaNE is a an efficient reinforcement learning (RL) framework for learning image-based robot manipulation tasks. It densifies sparse task rewards with exploration bonuses around demonstrations.

Prerequisites

• Python 3.10
• CUDA-compatible GPU (recommended)

Installation

1. Create and activate a conda environment:

conda create --name lane python=3.10.12
conda activate lane

2. Install dependencies:

pip install -r requirements.txt

Usage

Collecting Demonstrations

To collect and save task demonstrations, execute the following scripts:

python robosuite_utils/save_demo_lift.py
python robosuite_utils/save_demo_door.py
python robosuite_utils/save_demo_stack.py
python robosuite_utils/save_demo_pick_place_can.py

Training

To train the LaNE model on different tasks, use the following scripts:

bash scripts/lane_dino/robosuite_lift.sh
bash scripts/lane_dino/robosuite_door.sh
bash scripts/lane_dino/robosuite_stack.sh
bash scripts/lane_dino/robosuite_pick_place_can.sh

Scripts correponding to some baselines and ablation studies are also provided in the scripts folder.