Multi-Task Reinforcement Learning for Quadrotors

Why It Matters

A unified multitask policy cuts development time and hardware complexity while delivering high‑performance flight across diverse missions, accelerating commercial UAV adoption.

Key Takeaways

• Multitask RL yields a single quadrotor policy for diverse tasks.
• Separate task-specific and shared observations improve knowledge transfer.
• Shared encoder plus task-specific encoders boost sample efficiency.
• Single actor with multiple critics matches baseline performance across tasks.
• Real‑world tests confirm high‑speed stabilization, velocity tracking, racing.

Summary

The video introduces a multitask reinforcement‑learning framework that trains a single, generalist controller for quadrotors capable of handling stabilization, high‑speed racing, and velocity‑tracking commands. By partitioning sensor inputs into shared and task‑specific observations, the system feeds each through a common encoder and distinct task encoders before merging the embeddings for action prediction.

The architecture employs one actor network to output control signals while multiple critic networks provide task‑specific value estimates, enabling knowledge sharing across tasks and markedly improving sample efficiency relative to single‑task baselines. Experiments show the unified policy matches or exceeds performance of specialized controllers without additional training overhead.

Empirical results include successful high‑speed stabilization, accurate tracking of random velocity commands across a broad range, and agile maneuvering on a racetrack. The authors validate the approach on physical quadrotors in three separate real‑world scenarios, demonstrating robustness beyond simulation.

This work suggests that a single learned policy can replace multiple handcrafted controllers, simplifying deployment pipelines and accelerating the adoption of autonomous UAVs in varied operational contexts, from delivery to inspection.

Original Description

Reinforcement learning (RL) has shown great effectiveness in quadrotor control, enabling specialized policies to develop even human-champion-level performance in single-task scenarios. However, these specialized policies often struggle with novel tasks, requiring a complete retraining of the policy from scratch. To address this limitation, this paper presents a novel multi-task reinforcement learning (MTRL) framework tailored for quadrotor control, leveraging the shared physical dynamics of the platform to enhance sample efficiency and task performance. By employing a multi-critic architecture and shared task encoders, our framework facilitates knowledge transfer across tasks, enabling a single policy to execute diverse maneuvers, including high-speed stabilization, velocity tracking, and autonomous racing. Our experimental results, validated both in simulation and real-world scenarios, demonstrate that our framework outperforms baseline approaches in terms of sample efficiency and overall task performance.

Reference:

J. Xing, I. Geles, Y. Song, E. Aljalbout, and D. Scaramuzza,

"Multi-Task Reinforcement Learning for Quadrotors",

IEEE Robotics and Automation Letters RA-L

PDF: https://rpg.ifi.uzh.ch/docs/RAL25_Xing.pdf

More info on our research in Drone Racing:

https://rpg.ifi.uzh.ch/research_drone_racing.html

More info on our research in Agile Drone Flight:

https://rpg.ifi.uzh.ch/aggressive_flight.html

Affiliations:

J. Xing, I. Geles, Y. Song, E. Aljalbout, and D. Scaramuzza are with the Robotics and Perception Group, Dep. of Informatics, University of Zurich, and Dep. of Neuroinformatics, University of Zurich and ETH Zurich, Switzerland, https://rpg.ifi.uzh.ch/