TACO: Trajectory-Aware Controller Optimization for Quadrotors

TL;DR

This work tackles the problem of (trajectory-aware) tuning of quadrotor controllers, proposing Trajectory-Aware Controller Optimization (TACO) to adapt controller gains online based on the upcoming reference trajectory and current state. It introduces a predictive model $\hat{F}$ that maps $(g, o_n, \bar{\tau}_{n:n+H})$ to horizon-costs $c_{n:n+H}$ and uses receding-horizon optimization to adjust gains $g \in \mathbb{R}^8$, with an additional trajectory optimization capability through backpropagation of $\hat{F}$ to improve dynamic feasibility of polynomial splines. The approach is trained on a large, parallelized simulator dataset ($8\times10^6$ samples) and demonstrates that TACO outperforms static, hand-tuned baselines and is orders of magnitude faster than offline Bayesian optimization, enabling real-time deployment on hardware like the CrazyFlie; it also shows that online trajectory adaptation can substantially reduce tracking error. The work highlights practical benefits for real-time quadrotor control, while acknowledging sim-to-real gaps and proposing domain adaptation and faster adaptation as directions for future work.

Abstract

Controller performance in quadrotor trajectory tracking depends heavily on parameter tuning, yet standard approaches often rely on fixed, manually tuned parameters that sacrifice task-specific performance. We present Trajectory-Aware Controller Optimization (TACO), a framework that adapts controller parameters online based on the upcoming reference trajectory and current quadrotor state. TACO employs a learned predictive model and a lightweight optimization scheme to optimize controller gains in real time with respect to a broad class of trajectories, and can also be used to adapt trajectories to improve dynamic feasibility while respecting smoothness constraints. To enable large-scale training, we also introduce a parallelized quadrotor simulator supporting fast data collection on diverse trajectories. Experiments on a variety of trajectory types show that TACO outperforms conventional, static parameter tuning while operating orders of magnitude faster than black-box optimization baselines, enabling practical real-time deployment on a physical quadrotor. Furthermore, we show that adapting trajectories using TACO significantly reduces the tracking error obtained by the quadrotor.

Figures (6)

• Figure 1: We train a model to predict a quadrotor controller's performance based on the upcoming reference trajectory, and demonstrate that this model can be used to optimize the control parameters in real-time for different upcoming trajectory segments. We also use the same predictive model to adapt reference trajectories to increase dynamic feasibility.
• Figure 2: TACO's predictive model (left) and online gain optimization procedure (right)
• Figure 3: Trajectory adaptation using gradient backpropagation through the predictive model.
• Figure 4: 2D examples of the trajectory types included in our training dataset, interpolating between the same keypoints with the same average velocity.
• Figure 5: An example of tracking a challenging MinSnap trajectory using re-tuning with TACO (blue) versus the static oracle-tuned parameters (green) and adaptive oracle (orange). By retuning the control parameters online, TACO handles diverse maneuvers such going from a loop in altitude from $t=0:2$ to a gentle curve from $t=3:7$.