UZH researchers train quadrotor to throw payloads with 50% less error
Quadrotor learns to throw cable-suspended payloads accurately using RL simulation.
Researchers led by Davide Scaramuzza at the University of Zurich have developed a quadrotor capable of throwing cable-suspended payloads with unprecedented agility and accuracy. Their approach combines a high-fidelity analytical quadrotor model with a physics solver for complex rope-payload interactions, creating a hybrid simulation environment. A deep reinforcement learning policy trained in this environment was deployed zero-shot on real hardware, achieving 50% lower landing error and 30% faster throws compared to traditional model-based trajectory optimization methods. The system can also run on purely visual inputs, matching state-based accuracy without explicit position estimation.
The paper, titled 'Learning to Throw...', addresses a gap in dynamic aerial manipulation. While suspended-payload transport and traversal are well studied, targeted release under high dynamics remained challenging due to rope modeling difficulties. The hybrid simulation framework (exchanging forces at every step) enabled physically accurate training. Ablation studies confirmed the coupled simulation as the key enabler. The work promises applications in search-and-rescue and medical delivery where rapid, precise payload placement is critical. The simulator will be open-sourced to accelerate future research.
- Hybrid simulation framework couples quadrotor model with physics solver for rope dynamics
- RL policy reduces landing error by 50% and throw duration by 30% over model-based baselines
- Visual-only policy achieves comparable accuracy without explicit state estimation
Why It Matters
Enables drones to deliver payloads dynamically for rescue and medical drops with precision never before possible.