人工智能培训

搜索

人工智能教程:在仿真学习模拟QUADCOPTER上的高级导航指令之后(Following High-level Navigation Instructions

[复制链接]
baiselong 发表于 2018-6-4 08:14:12 | 显示全部楼层 |阅读模式
baiselong 2018-6-4 08:14:12 272 0 显示全部楼层
人工智能教程:在仿真学习模拟QUADCOPTER上的高级导航指令之后(Following High-level Navigation Instructions on a Simulated Quadcopter  with Imitation Learning)我们引入了一种方法,通过直接从图像,指令和姿态估计到连续的低级速度命令进行实时控制,从而跟踪高级导航指令。基于语义映射的网络(GSMN)是一种完全可微的神经网络体系结构,它通过在网络中结合针孔摄像机投影模型,在世界参考框架中构建一个明确的语义映射。存储在地图中的信息是从经验中学习的,而本地到世界的转换是明确计算的。我们使用DAggerFM来训练模型,DAgger是一种经过修改的DAgger变体,用于交易表格收敛性保证,以提高训练速度和内存使用。我们在真实的四坐标飞机模拟器上的虚拟环境中测试GSMN,并显示,结合一个明确的映射和接地模块可以使GSMN超越强大的神经基线,几乎达到专家级的策略性能。最后,我们分析学习的地图表示,并显示使用显式地图导致可解释的指令跟踪模型。
We introduce a method for following high-level navigation instructions bymapping directly from images, instructions and pose estimates to continuouslow-level velocity commands for real-time control.The Grounded SemanticMapping Network (GSMN) is a fully-differentiable neural network architecturethat builds an explicit semantic map in the world reference frame byincorporating a pinhole camera projection model within the network.Theinformation stored in the map is learned from experience, while thelocal-to-world transformation is computed explicitly.We train the model usingDAggerFM, a modified variant of DAgger that trades tabular convergenceguarantees for improved training speed and memory use.We test GSMN in virtualenvironments on a realistic quadcopter simulator and show that incorporating anexplicit mapping and grounding modules allows GSMN to outperform strong neuralbaselines and almost reach an expert policy performance.Finally, we analyzethe learned map representations and show that using an explicit map leads to aninterpretable instruction-following model.人工智能教程:在仿真学习模拟QUADCOPTER上的高级导航指令之后(Following High-level Navigation Instructions on a Simulated Quadcopter  with Imitation Learning) PcGapIqImI1QV4I4.jpg
URL地址:https://arxiv.org/abs/1806.00047     ----pdf下载地址:http://arxiv.org/pdf/1806.00047    ----人工智能教程:在仿真学习模拟QUADCOPTER上的高级导航指令之后(Following High-level Navigation Instructions on a Simulated Quadcopter  with Imitation Learning)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

baiselong当前离线
新手上路

查看:272 | 回复:0

快速回复 返回顶部 返回列表