About me

Hello, I’m a junior majoring in software engineering at Northwestern Polytechnical University. I have been a member of the Northwestern Polytechnic University V5++ robotics team (an undergraduate team focused on simulating 5 vs 5 soccer game, RoboMaster University AI Challenge and UAV) since April 2021.

Interested in T2S, Computer Vision, Robotics and Reinforcement Learning.

I have sufficient self-motivation, perseverance, optimism and cheerfulness, and hope to make my own contribution to scientific research. I’ll do my best!

Awards

2021&2022 Outstanding students of Northwestern Polytechnic University (First Class Scholarship)
2021&2022 Samsung scholarship
2022.05 International 3rd Prize of RoboMaster 2022 University AI Challenge (hosted by IEEE ICRA & DJI)
Technical report was rated A level (5/35) in RoboMaster 2022 University AI Challenge
2022.04 National Champion of 2021 China Robot Competition and RoboCup Robot World Cup FIRA Small Group
2021.10 National 1st Prize of the 23rd National Robot Championships
2021.08 National 1st Prize in Football Simulation Project of the 23rd China Robotics and Artificial Intelligence Competition
Several other national, provincial and ministerial awards……

Experience

2020.09 - present, Undergraduate, School of Software, Northwestern Polytechnic University, Xi’an, Shaanxi Province, China.
2022.07 - present, Vice Captain, Northwestern Polytechnic University V5++ robotics team
2021.04 - 2022.02, Research Intern, Audio, Speech and Language Processing Laboratory(ASPL) at Northwestern Polytechnic University

Projects

KWS Paper Code Reproduction

During an internship at ASLP, I reproduced the paper called “small-footprint keyword spotting using deep neural networks“, which is published in 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). The paper proposes a keyword spotting method called DeepKWS.

Here is my code.

My works:

Preprocess positive and negative audios containing four keywords, include fbank feature vector extraction and dataset labeling.
Build and train the DeepKWS original network.
Posterior Handling of the network outputs, include posterior smoothing and confidence score calculation.
Use Batch Normalization and Resnet to optimize the network to achieve better performance than the paper (but in fact sacrifice operation speed)

The ROC performance of the four-keyword spotting using DeepKWS in my work is similar to that of the paper, but there is still a gap.

net1 performance

According to the characteristics of Chinese audio speech in the data set, the network model was replaced, the size of the framing window was modified, etc., and the speed was sacrificed but the performance was better than that in the paper.

The ROC performances of the four-keyword spotting using origin network and modified network is shown in the figure below. The performance of the modified network is already better than in the paper.

net1&net2

RoboMaster AI Challenge

In the “Automatic Robot Shooting Confrontation” competition, participating teams need to use the official robot platform to perceive the environmental information of the battlefield, make independent decisions according to the situation on the field, and carry out motion planning and control.

Our Robot:

robot

My works:

I used Kalman filter to track the firing target, completed the automatic tracking system of the turret based on PID and feedforward control algorithm, and used the system identification method to better adjust the PID parameters. Our technical report for the 2022 season is rated A, ranked 5/35;
Responsible for the intelligent decision making part of the development, deployed two sets of behavior tree and reinforcement learning DDQN algorithms to control two robots autonomously against the enemy, while deploying gRPC for multi-robot communication, providing solutions to problems in decision making, navigation, control and other modules.