Self-Supervised Meta-Learning for All-Layer DNN-Based Adaptive Control with Stability Guarantees

Guanqi He, Yogita Choudhary, and Guanya Shi

Carnegie Mellon University

Abstract

A critical goal of adaptive control is enabling robots to rapidly adapt in dynamic environments. Recent studies have developed meta-learning-based adaptive control schemes that use meta-learning to extract nonlinear features (represented by Deep Neural Networks, or DNNs) from offline data, and apply adaptive control to update linear coefficients online. However, such schemes are fundamentally limited by the linear parameterization of uncertainties and do not fully unleash the capability of DNNs. This paper introduces a novel learning-based adaptive control framework that pretrains a DNN through self-supervised meta-learning (SSML) from offline trajectories and adapts the full DNN online via composite adaptation. In particular, the offline SSML stage leverages the time consistency in trajectory data to train the DNN to predict future disturbances from history in a self-supervised manner without environment condition labels. In the online stage, a control law and an adaptation law are carefully designed to update the full DNN while ensuring stability. Empirical results demonstrate that in challenging real-world quadrotor tracking problems under large dynamic wind disturbances, the proposed framework significantly outperforms (by 19-39%) various classic and learning-based adaptive control baselines.

Approach

(a) The goal of pretraining is to find an initial DNN parameter θ₀ such that, for all environment conditions w, θ₀ is close to the optimal DNN parameter θ*(w). (b) Pretraining θ₀ via self-supervised meta-learning (SSML) from offline trajectories. (c) Adapt the DNN parameter θ online using composite adaptive control.

Experiment Results

Experiments setting of quadrotor trajectory tracking under large dynamic wind conditions:

Disturbance prediction and tracking performance of each controller:

The experiment results show that the PID controller exhibits poor tracking performance due to the lack of wind disturbance feedforward compensation. The INDI controller estimates disturbances from IMU and motion capture system feedback, but applies a low-pass filter to disturbance estimation, leadning to phase delays and loss of high-frequency disturbance details, as shown in the second row of Fig. \ref{fig: dist-prediction}. For the vanilla-NN model, since the wind disturbance is dependent on the position of the drone, without meta-learning, it only learns an average disturbance prediction, lacking parameter flexibility in online adaptation.

In contrast, SSML-AC and SSML-AC-LL outperform all other baselines in both disturbance prediction and tracking accuracy, illustrating the importance of meta-learning in pretraining stage. Furthermore, SSML-AC further improves disturbance prediction accuracy and tracking performance, demonstrating the advantage of full-network adaptation. Repeated experiments show that both SSML-AC and SSML-AC-LL have lower variance than other baselines, highlighting the role of accurate disturbance representation in improving the consistency and robustness of online adaptation.

Video

draft6.mp4

Page updated

Google Sites

Report abuse