Online Adaptation of Learned Vehicle Dynamics Model with Meta-Learning Approach

TL;DR

The paper tackles online adaptation of learned vehicle dynamics near the limits of handling by applying Continual-MAML to a neural network model and evaluating its impact on a model-predictive path integral controller (MPPI) using a 1/10-scale TRIKart. The dynamics model is a two-hidden-layer neural network initialized from pre-training on a cement-floor dataset, with online adaptation to changing road conditions (rubber high-friction and foam low-friction). Results show that Continual-MAML outperforms a fixed model and a gradient-descent online learner in both inference accuracy and MPPI-based control performance, especially during surface transitions. This demonstrates effective continual, online adaptation for vehicle dynamics, enabling safer and more robust control in varying environments; future work includes automatic task-boundary detection and scaling to more challenging conditions.

Abstract

We represent a vehicle dynamics model for autonomous driving near the limits of handling via a multi-layer neural network. Online adaptation is desirable in order to address unseen environments. However, the model needs to adapt to new environments without forgetting previously encountered ones. In this study, we apply Continual-MAML to overcome this difficulty. It enables the model to adapt to the previously encountered environments quickly and efficiently by starting updates from optimized initial parameters. We evaluate the impact of online model adaptation with respect to inference performance and impact on control performance of a model predictive path integral (MPPI) controller using the TRIKart platform. The neural network was pre-trained using driving data collected in our test environment, and experiments for online adaptation were executed on multiple different road conditions not contained in the training data. Empirical results show that the model using Continual-MAML outperforms the fixed model and the model using gradient descent in test set loss and online tracking performance of MPPI.

Figures (6)

• Figure 1: TRIKart, a one-tenth scale-car which is used for our experiment. During experiments, electrical devices are covered by a plastic body and a bumper is mounted around the perimeter.
• Figure 2: Our test garage where 12 OptiTrack Prime 13 cameras are mounted on the ceiling. The cement floor in this garage is used for pre-training of a learned model.
• Figure 3: The oval course used for our experiments. Colors represent a trajectory cost, that is, the cost of a white area is 0 and a black area is 1. The gray is linearly interpolated between 0 and 1.
• Figure 4: Experimental road conditions for online adaptation. The right black mat is made with rubber which has high friction. The left gray mat is made with foam which has less friction than the rubber one.
• Figure 5: Enlarged view of inference loss for each model between 40 (s) to 60 (s). Each peak characterized by red dotted ellipses corresponds to turning on rubber mat. Especially in this region, differences between each method are significant. Gradient descent outperforms the fixed model and it shows effectiveness of adaptation. Continual-MAML outperforms gradient descent especially in turning on rubber mat and it implies Continual-MAML achieves quick and effective adaptation by starting update from optimal initial parameters when a road change is detected.