JuggleRL: Mastering Ball Juggling with a Quadrotor via Deep Reinforcement Learning
Overview
Highlights
- Zero-shot sim-to-real deployment, no real data for training.
- Calibrated dynamics + domain randomization to reduce sim-to-real gap.
- Lightweight Communication Protocol (LCP) for low-latency state streaming.
- Real-world performance: up to 462 hits (avg 311 across 10 trials).
This page hosts figures, demo videos, and links to paper & code.
Project Links
- Paper: https://arxiv.org/abs/2509.24892
- Training: https://github.com/thu-uav/JuggleRL_train
- ROS Pack: https://github.com/thu-uav/JuggleRL_rospack
- NatNet SDK: https://github.com/thu-uav/JuggleRL_NatNetSDK
Key Metrics
462
Max real-world hits
311
Avg hits (10 trials)
0
Real data for training
Full Demo on Bilibili
© 2025 JuggleRL Team · Hosted on Hugging Face Spaces