Task-Parameter Nexus: Task-Specific Parameter Learning for Model-Based Control

TL;DR

The paper introduces Task-Parameter Nexus (TPN), a learning-based framework that online-predicts near-optimal control parameters for model-based controllers to track arbitrary trajectories. It builds a trajectory bank spanning independent speed and curvature axes, uses batch-DiffTune to offline auto-tune expert parameters for each task, and trains a DNN to map trajectory tasks to parameter vectors, with constraints enforced by RAYEN to maintain stability. In quadrotor simulations, TPN achieves near-expert performance within the training bank and generalizes to unseen tasks inside the bank, with robust behavior on out-of-bank and trigonometrically parameterized tasks compared to untrained parameters. The findings demonstrate that combining a rich task bank, batch differentiable tuning, and supervised learning enables rapid, robust runtime adaptation of control parameters across diverse tasks, with potential to extend to other dynamic systems and to real-world flight tests.

Abstract

This paper presents the Task-Parameter Nexus (TPN), a learning-based approach for online determination of the (near-)optimal control parameters of model-based controllers (MBCs) for tracking tasks. In TPN, a deep neural network is introduced to predict the control parameters for any given tracking task at runtime, especially when optimal parameters for new tasks are not immediately available. To train this network, we constructed a trajectory bank with various speeds and curvatures that represent different motion characteristics. Then, for each trajectory in the bank, we auto-tune the optimal control parameters offline and use them as the corresponding ground truth. With this dataset, the TPN is trained by supervised learning. We evaluated the TPN on the quadrotor platform. In simulation experiments, it is shown that the TPN can predict near-optimal control parameters for a spectrum of tracking tasks, demonstrating its robust generalization capabilities to unseen tasks.

Figures (8)

• Figure 1: An illustration of the Task-Parameter Nexus when applied to a model-based controller.
• Figure 2: Illustration of the trajectory bank (consisting of trajectories under the hierarchy of category-parent-child-piece) and task batches for batch-DiffTune. Each trajectory piece is a task.
• Figure 3: Box plots of speed (a) and curvature (b) over the 12 categories in the trajectory bank.
• Figure 4: Trajectories ran by expert parameters (blue) vs untrained parameters (red) on three training tasks (black). Only three trajectories are shown for illustration purposes. Subfigures from left to right, top to bottom are from $\mathcal{S}_1 \mathcal{C}_1$, $\mathcal{S}_1 \mathcal{C}_4$, $\mathcal{S}_3 \mathcal{C}_1$, $\mathcal{S}_3 \mathcal{C}_4$.
• Figure 5: Illustration of parent 1 over all 12 categories. Each segment in between two dots has a duration of 2 s and represents one task. With increased speeds (from $\mathcal{S}_1$ to $\mathcal{S}_3$), the waypoints (dots) are wider spread, corresponding to faster speeds that pose challenges to translational motion tracking. With increased curvatures (from $\mathcal{C}_1$ to $\mathcal{C}_4$), the trajectories turn curvy and sinuous, corresponding to faster turns that pose challenges to rotational motion tracking.

Theorems & Definitions (2)

• remark 1
• remark 2