美国优达学城(Udacity)
纳米学位 Nanodegree
火星漫游车的搜索、挑选和返回 Mars Rover Search, Pick, and Return
对火星漫游车进行编程，使其在模拟环境中自主导航并绘制环境地图，同时搜索感兴趣的岩石样本。在绘制了 40% 的地图后，它达到了 60% 以上的保真度。
Program a rover to navigate autonomously in simulation and create a map of its environment, while searching for rock samples of interest. It achieved over 60% fidelity after mapping 40% of the map.
仿真库卡KR210机器人的物体拾放 Pick & Place with Kuka KR210 robot
在Gazebo仿真中执行库卡 KR210 的逆运动学，并完成 10 次拾取和放置任务。
Perform inverse kinematics for the Kuka KR210 in Gazebo simulation and complete pick and place tasks 10 times.
仿真PR2机器人的3D物体识别 3D Object Recognition with PR2 robot
使用PR2机器人在Gazebo仿真里进行一系列物体识别。
Perform sequential object recognition using the PR2 robot in Gazebo.
机器人视觉识别 Robotic Inference
在该机器人视觉识别项目中，我提出了一个包含三种不同物体类别的物体分类问题，并通过基于自收集数据集训练标准预训练网络来解决该问题。为了进一步证明预训练识别模型的能力，我在训练和测试过程中还考虑了每个对象不同的颜色、不同的摆放和不同的背景。在三个不同的物体类别的情况下，识别的准确率达到了 90% 以上。未来的工作包括使用不同的预训练网络（如 AlexNet 或更新的网络）来比较其性能，以及使用彩色图像作为输入来进一步改进识别模型。
In this robotic inference project, an object classification problem with three different object classes is raised and solved by training a standard pre-trained network based on a self-collected dataset. For each object, different object colors, different object poses, and different backgrounds are also considered in the training and testing process to further prove the ability of the pretrained inference model. The inference accuracy achieves over 90% on three different object classes. Future work includes using different pretrained networks like AlexNet or the newer ones to compare the performances, as well as using color images as input to further improve the inference model.
全卷积网络(FCN)四旋翼飞行器行人追踪 Quad-copter Pedestrian Follow using FCN
搭建并训练全卷积网络(FCN)，以从模拟四旋翼飞行器的图像中找到特定的人
Build and Train a Fully Convolutional Network (FCN) to find a specific person in images from a simulated quad-copter
机器人定位 Where Am I?
我在本项目的模拟定位任务中应用了自适应蒙特卡洛定位。一个基准机器人和一个定制机器人在豺狼赛跑地图上执行了任务，机器人被放置在地图中间，需要导航到目标位置和方向。根据 ROS wiki 页面将参数调整到合适的值后，两个机器人都能在大多数情况下以合理的路径在合理的时间内到达最终目标。未来的工作包括创建一个完全不同的机器人，配备多个不同的传感器，并添加更多参数，以实现更好的性能。另一项未来工作是在真正的移动机器人上安装 AMCL，并在实验室环境中测试其能力。
In this project, an adaptive Monte Carlo localization is implemented in a simulated localization task. A benchmark and a customized robot have performed the task on a jackal race map, where the robots are put in the middle of the map and are required to navigate to the goal position and orientation. Both robots are able to reach the final goal in a reasonable path in most scenarios within a reasonable time after tuning the parameters to appropriate values based on the ROS wiki page. Future works include creation of a completely different robot with multiple different sensors as well as adding more parameters in order to achieve better performances. Another future work is to setup the AMCL on a real mobile robot and test its ability in a lab environment.
环境绘制机器人 Map My World Robot
本项目的目标是在机器人上实施映射算法，生成两个模拟世界的可识别地图。为进行测试，我在 Gazebo中创建了一个定制的咖啡馆世界。使用安装了RGB-D摄像头和激光传感器的双轮机器人执行绘图任务。使用ROS、Gazebo和基于外观的实时绘图（RTAB-Map）创建了两个环境的二维占位网格和三维八维地图。在机器人探索了两个环境后，地图被准确生成。未来的工作还包括在机器人实验室环境中安装 RGB-D 摄像头和激光雷达的真实先锋机器人上应用RTAB-Map。
In this project, the goal is the implementation of the mapping algorithm on a robot to generate recognizable maps of two simulated worlds. A customized cafe world is created in Gazebo for testing. A two-wheel robot with a RGB-D camera and lasor sensor installed is used to perform the mapping task. A 2D occupancy grid and a 3D octomap of two environments are created using ROS, Gazebo, and Real-Time Apperarance-Based Mapping (RTAB-Map). After the robot explored both environments, the maps were generated accurately. Future work also includes the implementation of RTAB-Map on a real Pioneer robot with a RGB-D camera and LiDAR installed in a robotic lab enviroment.
机械臂的深度强化学习(Deep RL) Deep Reinforcement Learning (Deep RL) Arm Manipulation
本项目的目标是创建一个DQN代理并定义奖励功能，以教会机械臂实现两个主要目标：
The goal of this project is to create a DQN agent and define reward functions to teach a robotic arm to carry out two primary objectives:
1. 让机械臂的任何部分接触到感兴趣的物体，准确率至少达到 90%
Have any part of the robot arm touch the object of interest, with at least an 90% accuracy
2. 只让机械臂的抓手基座接触到物体，准确率至少达到 80%
Have only the gripper base of the robot arm touch the object, with at least an 80% accuracy
该项目在Nvidia Jetson TX2上进行训练。奖励和超参数会在多次试验后进行修改
The project is trained on Nvidia Jetson TX2. The rewards and hyperparameters are modified after several trials.
最终目标1在训练完593 epochs 后达到了93.53%准确率（图1）、目标2在1900 epochs 后达到了80.03%准确率（图2）
The training achieves 93.53% for Objective 1 with 593 epochs (Fig 1) and 80.03% for Objective 2 with 1900 epochs (Fig 2).
家庭服务机器人 Home Service Robot
为家庭服务机器人编程，使其能够自主绘制环境地图，并导航取送物品
Programmed a home service robot that can autonomously map an environment and navigate to pick up and deliver objects
神经网络预测自行车日租赁量 Daily Bike Rental Ridership Prediction with Neural Network
在这个项目中，我搭建了一个神经网络来预测自行车租赁的日骑行量。
In this project, I built a neural network to predict daily bike rental ridership.
卷积神经网络(CNN)图像分类 Image Classification using CNN
在这个项目中，我对CIFAR-10数据集中的图像进行了分类。该数据集包括飞机、狗、猫和其他物体。我对图像进行了预处理，然后在所有样本上训练了一个卷积神经网络(CNN)。对图像进行了归一化处理，并对标签进行了one-hot编码。我构建了卷积层、最大池化层、剔除层和全连接层。我的神经网络预测样本图像的测试准确率为62.9%。
In this project, I classified images from the CIFAR-10 dataset. The dataset consists of airplanes, dogs, cats, and other objects. I preprocessed the images and then trained a convolutional neural network (CNN) on all the samples. The images were normalized, and the labels were one-hot encoded. I built convolutional, max pooling, dropout, and fully connected layers. My neural network predicts the sample images with a testing accuracy of 62.9%.
递归神经网络(RNN)电视脚本生成 TV Script Generation using RNN
在这个项目中，我使用递归神经网络 (RNN) 生成了自己的《辛普森一家》电视脚本。我使用了《辛普森一家》数据集中 27 季的部分脚本。我构建的神经网络为莫氏酒馆的一个场景生成了一个新的电视脚本。
In this project, l generated my own Simpsons TV scripts using recurrent neural networks (RNNs). I used part of the Simpsons dataset of scripts from 27 seasons. The Neural Network I built generates a new TV script for a scene at Moe's Tavern.
递归神经网络(RNN)语言翻译 Language Translation using RNN
在这个项目中，我在一个英语和法语句子数据集上训练了一个sequence-to-sequence模型，该模型可以将新句子从英语翻译成法语。
In this project, I trained a sequence-to-sequence model on a dataset of English and French sentences that can translate new sentences from English to French.
生成对抗网络(GAN)人脸生成 Face Generation using GAN
在这个项目中，我使用生成对抗网络（GAN）生成新的人脸图像。
In this project, I used generative adversarial networks (GANs) to generate new images of faces.
