Projects

SOOT Morph Investigator

Intersection project. Introduce CV into atmospheric particle identification. Built an automatic program for batch processing electro-microscope images.

Period: Oct 2019 - Jan 2020

paper

Pretrained Model for Bounding Gait

My first DeepRL project. Extract reflections from data trajectories to boost RL optimization.

Period: Dec 2019 - Aug 2020

paper

Hybrid Bipedal Locomotion

DRL with humanoid locomotion. Tried to reduce NN policy’s stress by combining Raibert’s controller.

Period: Feb 2022 - Aug 2022

paper

Learned Humanoid Locomotion

Trying to train robust locomotion policies with DRL

Period: Jan 2021 -

Minimal Learned Quadrupedal Locomotion

First project during A*STAR attachment.

Built a minimal RaiSim training environment that produces robust trotting policies that can work on real robots.

Aiming to provide a baseline for following privilege training implementations.

Period: Jul 2022 - Aug 2022

Regularized RMA

An improved implementation of RMA

Beside reappearing the teacher-student privilege training schedule, added specialized gait reward with inspiration from Peroidic Reward and control commands.

Training curriculum is also modified into a simpler version.

Trained trotting, bounding, galloping and fall recovery policies.

Period: Oct 2022 - Jan 2023

paper

SDK development for Wukong-IV humanoid

Wukong-IV is a humanoid robot designed and made by Robotics & Machine Intelligence Lab of Zhejiang University. It is 1.4 m tall, weighs 45 kg, and is actuated by 21 electric motor joints. The SDK was formalized in this project, as the preparation for subsequent developments.

Period: Aug 2023 - Oct 2023

Learned Robust Humanoid Locomotion

A humanoid implementation based on DreamWaQ

Trained robust walking policies that can traverse stairs, slopes and outdoor terrains. The policy also can endure kicking.

Related videos (on Bilibili.com):

Period: Oct 2023 - Mar 2024

paper