Whole-Body Imitation by Humanoid Robots and Task-Oriented Teleoperation
Aug 2016 - May 2018 Bionic Intelligent Robot Lab Advised by Prof. Zhijun Zhang
We developed a humanoid robot's whole-body imitation system enabling the imitation of head motions, arm motions, lower-limb motions, hand motions and locomotion and not requiring any ancillary handheld or wearable devices or any additional audio or gesture-based instructions. A task requiring standing on one foot, grasping and walking can be performed by teleoperation using this system in real time. As an extension to this system, we trained an offline imitation learning model using GMM and GMR (originally proposed by Calinon et al.). Most of this work was done by Yaru.
Active Hierarchical Imitation and Reinforcement Learning
Fall 2019 and Fall 2020
In this work, we developed an Active Hierarchical Imitation and Reinforcement Learning (AHIRL) algorithm which can work in continuous spaces. We used DDPG to pre-train the low-level controller, and employed DAgger to train our meta-controller from human demonstrations. What is more, several active learning methods were designed and used for improving the sample efficiency of the hierarchical learning.
Formation Control and Collision Avoidance using Multi-Agent Policy Gradient
Fall 2019 Georgia Tech ECE 6563
In this project, we designed formation control and collision avoidance tasks with shaped reward functions, and explored two policy gradient methods, DDPG and MADDPG, which can be applied in the multi-agent environment.
DietMate - A Multimodal Diet Monitoring System
Summer 2018 AICPS Lab, UC Irvine Advised by Prof. Al Faruque
This project is about monitoring people’s eating and drinking behaviors using the data obtained from multiple sensors, to help people maintain balanced diets and healthy lifestyles. In the project, I designed algorithms that employed signal processing techniques and machine learning algorithms to process the data, extracted features from the data and estimated the behaviors of eating and drinking.
Analysis of Influencing Factors on Humanoid Robots' Emotion Expressions by Body Language
January 2018 - April 2018 Bionic Intelligent Robot Lab Advised by Prof. Zhijun Zhang
In this project, we explored humanoid robots’ capabilities of expressing emotions by body language and discussed the factors that may influence the emotion expression through questionnaire surveys and the statistic analysis. The research is carried out on the Nao robot and the results show that the expression of emotion is affected by the ambiguity of the body language and the joint limits of the robot.
Gesture - Determined Dynamical Schemes for Motion Planning of Humanoid Robot Arms
January 2018 - May 2018 Zhike Intelligence Advised by Prof. Zhijun Zhang
In this project, we explored and designed methods for Gesture - determined Dynamical Schemes for Motion Planning of the redundant manipulator, which can be applied to humanoid robots. In this way, the humanoid robot can not only perform end-effector tasks, but also can act following a sequence of specific gestures. Then we improved the robustness of the gesture-based dynamical scheme (avoided the joint limit) by modifying the gesture-based dynamical function which formulates the margins of the quadratic programming problem.