The accumulation of errors in the positioning of intelligent vehicle can pose a threat to the safety of vehicle driving. To eliminate dependence on high-precision positioning information, deep learning-based autonomous guidance technology for intelligent vehicles is gradually emerging. However, typical deep learning models can only be adapted to the task environment they are trained for. In this paper, we address the problem of obtaining more powerful student models based on multiple trained submodels to adapt to various task environments. We propose an knowledge amalgamation method based on Mixture-of-Experts (MoE) to construct asymptotically growable autonomous guidance models for intelligent vehicles. As the number of submodels increases, the autonomous guidance model expands in a manner analogous to human learning processes. The method in this paper is validated in CARLA simulation environments, demonstrating better performance and adaptability compared with previous methods.
A Knowledge Amalgamation Method for Autonomous Guidance of Intelligent Vehicles
2024-11-08
1050767 byte
Conference paper
Electronic Resource
English