---
## Abstract
**FLARE** is a reinforcement learning (RL) framework designed to tackle the formidable challenge of agile flight for quadrotor cable-suspended payload systems. Due to the underactuated, highly nonlinear, and hybrid dynamics of such systems, traditional methods often struggle.
In this work, we present a method that:
- **Directly learns** an agile navigation policy from high-fidelity simulation.
- Outperforms state-of-the-art optimization-based approaches (Impactor) by a **3x speedup** in gate traversal.
- Achieves successful **zero-shot sim-to-real transfer**, demonstrating remarkable agility and safety in real-world experiments.
## Installation
This repository depends on [GenesisDroneEnv](https://github.com/KafuuChikai/GenesisDroneEnv) and the official [Genesis](https://github.com/Genesis-Embodied-AI/Genesis) simulator.
⚠️ **Important**: To ensure stable reproducibility, we **strongly recommend** using the specific commit version of Genesis that this project was developed on.
```bash
# 1. Clone Genesis and checkout the specific commit
git clone https://github.com/Genesis-Embodied-AI/Genesis.git
cd Genesis
git checkout 382cf4ca12c0c142adcf2fa7675eef65caf0c661
pip install -e .
# 2. Install rsl_rl
git clone https://github.com/leggedrobotics/rsl_rl
cd rsl_rl
git checkout v1.0.2
pip install -e .
# 3. Clone Flare
git clone https://github.com/BEI11HAI/Flare.git
cd Flare
```
## Usage
We provide three challenging scenarios for validation: **Agile Waypoint Passing**, **Payload Targeting**, and **Agile Gate Traversal**.
Below are the instructions to evaluate pre-trained models or train your own policies for **Scenario I: Agile Waypoint Passing**. (Instructions for other scenarios will be updated soon).
### 1. Evaluate Pre-trained Policy
You can directly evaluate the pre-trained model provided in `logs/s1_waypoint_passing`.
```bash
python waypoint_passing_eval.py
# Add --record to save a video
# python waypoint_passing_eval.py --record
```
Then you will see:
