Progress Update | 进度汇总 2019.08.01-2019.08.28


#1

What We Have Worked On 2019.08.01-2019.08.28

Major Bug Fixed

We have fixed a bug related to calculation of SLAM rotation angle. Now the SLAM algorithms allow UAVs to have more stable performance in long-distance flight. We have seen a notable improvement in performance in the simulator.

GAAS now supports AirSim!

We switched to AirSim to collect new image data to test the performance of the algorithms in a more realistic simulation environment. We have also published a new case study to provide a step-by-step configuration of GAAS with AirSim.

Added New Global Optimization Module

We have added a new global optimization module to optimize the position and altitude of UAVs globally. When the flight control or SLAM is being rebooted, the position and altitude information can ensure the stability of position estimation.

What We Are Working On

We are looking into the viability of a new neural network that will potentially improve the performance of existing depth estimation algorithms, which has limited performance.

We are updating the mapping algorithms. Existing Octomap algorithm is one of the representations of Occupancy, but a lot of route navigation algorithms require information other than Occupancy.

We are enhancing the multi-sensor fusion algorithm to improve system tolerance.

What We Will Work On Next

  • Improve multi-sensor fusion algorithm
  • Use voxblox to replace Octomap algorithm for mapping projection
  • Establish dataset based on AirSim to train neural networks to optimize depth estimation.

我们做了什么 2019.08.01-2019.08.28

在过去的一个月里,我们修复了 SLAM 的旋转角度计算错误 Bug,现在的 SLAM 能够让无人机在长距离工作时飞得更加稳定,在模拟器中的表现有了很大提升。

为了更好的模拟真实环境,我们使用了 AirSim 搭建了一些场景,用来测试无人机在更加真实的模拟环境中的飞行表现,以及采集数据用于测试新的算法,我们写了一个新的教程帮助其他人配置 AirSim 以及在 AirSim 中使用 GAAS。

我们为多传感器融合加入了全局优化模块,用以全局优化飞行器的位姿,并且保存这些信息以便在飞控或SLAM状态重置/重启动时仍然有稳定的位置估计。

我们最近关心什么

由于传统双目算法在深度计算上效果有限,我们正在研究神经网络算法的可行性。

更新地图映射算法,目前我们采用的 Octomap 算法是一种 Occupancy 表示方法,但是很多路径规划算法只需要占用信息是不够的。

我们也要完善多传感器融合的算法,提高系统整体的冗余度和容错性。

我们接下来要做什么

  • 完善多传感器融合部分

  • 使用 voxblox 代替原有的 Octomap 算法用于地图映射。

  • 基于 AirSim 建立数据集,并训练神经网络优化深度计算部分。