# A Learning-based Quadcopter Controller with Extreme Adaptation

**Authors:** Dingqi Zhang, Antonio Loquercio, Jerry Tang, Ting-Hao Wang, Jitendra Malik, Mark W. Mueller  
**arXiv:** [2409.12949](https://arxiv.org/abs/2409.12949)  
**Date:** 2024-09-19  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2409.12949v1/

## Key points

- Introduction of a novel learning-based low-level controller for quadcopters
- Controller designed to adaptively control quadcopters with variations in mass, size, and actuator capabilities
- Approach combines imitation learning and reinforcement learning for fast-adapting and general control framework
- Controller estimates latent representation of vehicle's system parameters from sensor-action history for swift adaptation to diverse dynamics
- Extensive evaluations in simulation demonstrate ability to generalize to unseen quadcopter parameters with adaptation range up to 16 times broader than training set
- Real-world tests validate effectiveness on quadcopters with mass differences of 3.7 times and propeller constants varying by more than 100 times
- Acknowledgments to various sources including support from Hong Kong Center for Logistics Robotics, DARPA TIAMAT program, UC Berkeley Dept. of Mechanical Engineering Graduate Division Block Grant, ONR MURI award N00014-21-1-2801
- Special thanks extended to Ruiqi Zhang and Teaya Yang for assistance with experiments
- Experimental testbed at HiPeRLab noted as collaborative effort with contributions from many individuals listed on their website hiperlab.berkeley.edu/members/
- Showcases potential of extreme adaptation in simplifying design processes and enhancing reliability for autonomous drone operations in unpredictable environments

## Summary

This paper introduces a novel learning-based low-level controller for quadcopters. The controller is designed to adaptively control quadcopters with significant variations in mass, size, and actuator capabilities. By combining imitation learning and reinforcement learning, the approach creates a fast-adapting and general control framework that eliminates the need for precise model estimation or manual tuning. The controller estimates a latent representation of the vehicle's system parameters from sensor-action history, allowing it to swiftly adapt to diverse dynamics. Extensive evaluations in simulation demonstrate its ability to generalize to unseen quadcopter parameters, with an adaptation range up to 16 times broader than the training set. Real-world tests further validate its effectiveness on quadcopters with mass differences of 3.7 times and propeller constants varying by more than 100 times. Acknowledgments go out to various sources including support from the Hong Kong Center for Logistics Robotics, DARPA Transfer from Imprecise and Abstract Models to Autonomous Technologies (TIAMAT) program, Graduate Division Block Grant of the Dept. of Mechanical Engineering at UC Berkeley, and ONR MURI award N00014-21-1-2801. Special thanks are extended to Ruiqi Zhang and Teaya Yang for their assistance with experiments. The experimental testbed at HiPeRLab is noted as being a collaborative effort with contributions from many individuals listed on their website hiperlab.berkeley.edu/members/. This work showcases the potential of extreme adaptation in simplifying design processes and enhancing reliability for autonomous drone operations in unpredictable environments. For more information, viewers can access the video demonstration and code repository at https://github.com/muellerlab/xadapt_ctrl.