Abstract
This paper introduces a streamlined, data-driven method to identify a quadrotor's physical parameters—including the often-ignored motor delays—using just one minute of flight data and no specialized hardware. By accurately modeling these "hidden" dynamics, the researchers enable high-performance controllers and Reinforcement Learning (RL) policies to transition seamlessly from simulation to challenging real-world environments.
Introduction
Modern quadrotor control is moving away from traditional PID loops toward more sophisticated methods like Model Predictive Control (MPC) and Reinforcement Learning (RL). While these methods offer superior performance, they are "model-hungry": they require an extremely accurate mathematical description of the drone's physics to work.
Traditionally, acquiring parameters such as the inertia matrix or thrust curves is a significant challenge, often requiring drone disassembly or expensive external equipment like thrust stands. Furthermore, most models ignore motor delay—the slight lag between a command being sent and the rotors reaching the desired speed. Because standard drones lack sensors to measure actual RPMs, this delay is a "latent" or hidden variable that is notoriously difficult to estimate. Ignoring it leads to "noisy" data that frequently causes sim-to-real transfer to fail.
Methodology
The proposed "gray-box" approach combines physical first principles with data-driven optimization. The pipeline requires only proprioceptive data (onboard accelerometer and gyroscope readings) and basic measurements like total mass and rotor positions.
Motor Delay Estimation
The authors model the motors as a first-order system defined by a time constant ($T_m$). Since RPMs are not directly observable, they employ a Maximum A Posteriori (MAP) estimation. By performing a one-dimensional sweep over possible delay values, they identify the constant that minimizes the error between predicted and observed accelerations.
Thrust and Inertia Identification
Once the delay is identified, the actual motor speeds at any timestamp can be estimated using an Exponential Moving Average (EMA) filter. This transforms complex physics equations into a linear "Ax = b" problem, solvable via standard least-squares regression. To resolve the mathematical coupling between yaw inertia and torque coefficients, the authors leverage a scaling law ($I \propto \text{scale}^5$) observed across various drone models, establishing a reliable linear relationship between the axes.
Results and Key Insights
The method was validated on platforms ranging from the 27g Crazyflie 2.1 to a 3.35kg custom heavy-lifter.
The "Clean Data" Revelation
A pivotal finding of this work is the impact of motor delay modeling on data quality. Without accounting for delay, the relationship between motor commands and actual thrust appears as a scattered distribution. Factoring in the identified delay "cleans" this data into a clear, predictable curve, which is essential for training robust models.
Performance and Robustness
- Accuracy: For the Crazyflie, the identified motor delay ($T_m = 0.072s$) closely matched manufacturer step-response data ($\approx 0.073s$).
- Efficiency: Parameters for a large quadrotor were identified from just 73 seconds of flight data.
- Robustness: Using the identified model, an RL policy trained in simulation was deployed on a 3.35kg drone, successfully tracking trajectories in outdoor conditions with gusty winds up to 8 m/s without manual tuning.
Conclusion
This research demonstrates that accurate system identification can be achieved rapidly using onboard sensors alone. By treating motor speed as a recursive filter of command setpoints, the need for expensive RPM sensors is bypassed. The authors have released a web-based tool (sysid.tools) to allow the community to generate parameters from flight logs. Future work will address asymmetric motor dynamics and extending the tool to even more diverse aerial platforms.
Summary updated