Search Results

Deep Deterministic Policy Gradient (DDPG) is one of the Deep Reinforcement Learning algorithms. Because of the well perform in continuous motion control, DDPG algorithm is applied in the field of self-driving. Regarding the problems of the instability of DDPG algorithm during training and low training efficiency and slow convergence rate. An improved DDPG algorithm based on segmented experience replay is presented. On the basis of the DDPG algorithm, the segmented experience replay select the training experience by the importance according to the training progress to improve the training efficiency and stability of the training model. The algorithm was tested in an open source 3D car racing simulator called TORCS. The simulation results demonstrate the training stability is significantly improved compared with the DDPG algorithm and the DQN algorithm, and the average return is about 46% higher than the DDPG algorithm and about 55% higher than the DQN algorithm.

For intelligent vehicles, a robust perception system relies on training datasets with a large variety of scenes. The architecture of federated learning allows for efficient collaborative model iteration while ensuring privacy and security by leveraging data from multiple parties. However, the local data from different participants is often not independent and identically distributed, significantly affecting the training effectiveness of autonomous driving perception models in the context of federated learning. Unlike the well-studied issues of label distribution discrepancies in previous work, we focus on the challenges posed by scene heterogeneity in the context of federated learning for intelligent vehicles and the inadequacy of a single scene for training multi-task perception models. In this paper, we propose a federated learning-based perception model training system.

In the process of designing the aerodynamic kit for Formula SAE racing cars, there is a lot of repetitive work and low efficiency in optimizing parameters such as wing angle of attack and chord length. Moreover, the optimization of these parameters in past designs heavily relied on design experience and it's difficult to achieve the optimal solution through theoretical calculations. By establishing a parametric model in CAD software and integrating it with CFD software, we can automatically modify model parameters, run a large number of simulations, and analyze the simulation results using statistical methods. After multiple iterations, we achieve fully automatic parameter optimization and obtain higher negative lift. At the same time, the simulation process is optimized, and simulations are run based on GPUs, resulting in a significant increase in simulation speed compared to the original.