Sim2real: Issues in transferring autonomous driving model from simulation to real world
Sim2real: Issues in transferring autonomous driving model from simulation to real world
复制标题
Sim2real：自动驾驶模型从模拟转移到现实世界的问题
DOI：
10.1109/southeastcon48659.2022.9764110
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
J. Hereford
中科院分区：
文献类型：
--
作者：
Jacob Revell;Dominic Welch;J. Hereford
在这项研究中，我们调查的问题，sim 2真实的性能，这是发生的问题时，从训练模型在模拟不结转可比的真实的世界的结果。我们使用DeepRacer-for-Cloud软件训练Amazon Web Services DeepRacer汽车，以在模拟的椭圆形轨道上行驶。然后，我们在真实的赛道上测试DeepRacer赛车。我们在训练过程中考虑不同的动作空间，不同的奖励函数和不同的熵值（探索），看看哪一个能提供最好的真实世界性能。我们的研究结果表明，更简单的行动空间，更简单的奖励和更小的熵给最好的sim 2 real性能。
In this research we investigate the issue of sim2real performance, which is the problem that occurs when results from training a model in simulation do not carry over to comparable real world results. We train an Amazon Web Services DeepRacer car using DeepRacer-for-Cloud software to navigate a simulated oval track. We then test the DeepRacer car on a real-world track. We consider different action spaces, different reward functions and different values of entropy (exploration) during training to see which gives the best real-world performance. Our results show that the simpler action space, simpler reward and smaller entropy give the best sim2real performance.