I am a graduate student in Shanghai Jiao Tong University and studying as a master of robotics. Now I am interested in 5R: Reinforcement learning techniques , Reasoning , and Representation learning used in Robotics , invariably fascinated by Real intelligence .
When I was an undergraduate student of mechanical engineering, the thought of intelligent robots deeply attracted me. Therefore, I chose robotics intelligent control as my research direction. I imagine that one day every procedures of robots design can reduce human manipulation as much as possible. More frantically, giving the real intelligence of thinking and decision to robots is significant to humankind for both lifestyle changing and race evolution.
Besides dreaming about being a creator, human life is still colorful and romantic for me. Classical music is my favorite, especially Piano Concerto No. 1 by Tchaikovsky and I often go to concerts on weekends. In addition, traveling is also a kind of routine, I enjoy the different lifestyles in each new city and eager to explore that. I think the experience is the meaning of life and what we need is just to take a step and to trial boldly.
Robotics is a system engineering, so that we need to know not only software knowledge but also how to design and build a robot from jumbled hardware parts. Benefit from the style of the lab in my master period, the products we built are all accorded to application and the skills of packaging programs are also developed.
I got my bachelor diploma in 2018 from Southwest Jiaotong University and will be graduated as a master of robotics in 2021 from Shanghai Jiao Tong University. During this period, I have been participated several internships and summer camps.
Engaged with Huawei robotics research on enhancing the active object navigation of domestic robot employing Reasoning and Representation learning.
At the first year, I have studied some basic Control Theory lessons such as Modern Control and Intelligent Control. Besides, DSP (Digital Signal Processing), Matrix Theory, Optimization and some Robotics lessons were also compulsory.
Deecamp, one of China's largest AI training camps, initiated by renowned artificial intelligence expert Kaifu Lee who is also the CEO of Sinovation Ventures. At the beginning of the camp, we were able to listen the speech of plenty of AI experts from both academia and industry and have a chance to talk to them face to face. It is really a period of brainstorming. Then we were separated into many groups and tried to use AI techniques to solve practical problem. For me, as a captain of Shanghai 46th team, I led several graduates students from some top schools to deploy Reinforcement Learning technique on intelligent traffic control system. After about one month team working, our project achieved a better performance and can reduce the congestion. Fortunately, we were rewarded the Best Technology Award.
Learning how to use the classical KUKA robot manipulator first and program it for implementing specific grasping and shifting.
In the four years of my undergraduate life, I have learnt about three main parts: Mechanical Engineering, Electronic Engineering and Control Theory.
Here contains some open source projects of my own and my lab.
Note that some of them might be private currently.
2020.07-present
(Submitted to RA-L) ReVoLT: Relational Reasoning and Voronoi Local graph planning for Target-driven navigation
Engaged with Huawei robotics research on enhancing the active object navigation of domestic robot
employing
reasoning and representation learning. Improved the robot target-driven navigation task by abstracting
its planning procedure into a bandit problem with the
proposed reasoning method based on structured prior knowledge, expecting to realize an intelligent
reasoning
exploration. Built a representation graph rendering the observed scenes and exploited it as an
environment model; Adopted this
method to replace the explicit SLAM mapping and simple function approximation, delivering better
generalization,
robustness and flexibility for scene updates. The research article will be submitted to the CVPR2021
conference for review.
2020.02-present
(To be submitted to CoRL2021) Efficient
reinforcement learning control for continuum robots based on Inexplicit Prior Knowledge
Led the research project in collaboration with Shanghai Ruijin Hospital, exploring the efficient
reinforcement learning
(RL) control for continuum robots based on inexplicit prior knowledge. Proposed a new and data-efficient
model-based RL framework that integrates inexplicit prior knowledge (IPK) using
Kalman filter and can be directly deployed to the robot without simulation. The research article will be
submitted to the IROS2021 conference for review.
2019.07-2019.12
(Published in EAAI) Learning Scalable Multi-Agent Coordination by Spatial Differentiation
for Traffic Signal Control
Led the team and applied the concept of multi-agent reinforcement learning in building an intelligent
traffic control
system. Developed, combined with Attention Mechanism, a scalable multi-agent coordination by spatial
differentiation for
traffic signal control, managing to relieve the congestion and make decisions based on the analysis of
comprehensive
traffic conditions. Rewarded the Best Technology Award in Deecamp and published the research article in
EAAI journal.
2019.06-2020.03
(Published in SAE) A Multi-Modal States based Vehicle
Descriptor and Dilated Convolutional Social Pooling for Vehicle Trajectory Prediction
Proposed a multi-modal vehicle description and dilated social pooling based vehicle trajectory
prediction, helping the
autonomous vehicles judge the coming cars’ intention of changing lanes. Achieved better accuracy in
public data set than that of the SOTA algorithms. The research article was published in SAE
International Conference.
2019.07-2019.12
(Published in Mechanical and electronic engineering) A novel cRes-GAN algorithm for thyroid node
detection and classification
Developed a novel cRes-GAN algorithm for detecting and classifying thyroid nodes. Designed the cRes-GAN
algorithm based on 1501 original samples in the DICOM format, significantly expanding the
data conditions and increasing the diagnosis accuracy to 92.2%. The proposed method was adopted by
Shanghai Ruijin Hospital in clinical treatment as an auxiliary diagnosis method;
the research article was published in Mechanical Engineering and Technology.
2019.01-2019.07
Robomaster 2019 Global Finals Silver Medal Winner
Participated in the competition and worked with the teammates on designing, assembling and controlling
four kinds of
robots. The designed robots were tested fortheir electronic control level in dual meet; obtained the
Silver Medal Winner in the
national finals.
2018.09-2019.01
DDPG control for an automatic transmission robot
Designed a DDPG control for an automatic transmission robot for Pan Asia Technical Automotive Center
(PATAC);
Developed an autopilot for tracking the vehicle’s speed using robotics techniques, aiming to meet the
WLTC standard implemented by automakers in the emission test. Utilized Reinforcement Learning’s deep
deterministic policy gradient (DDPG) to assist the robot to control the speed of
a real-world vehicle following the WLTC curve.
jjliu@mae.cuhk.edu.hk, skylark0924@gmail.com
Zhihu Blog: https://www.zhihu.com/people/JunjiaLiu