cv

Ph.D. (Candidate)

University of Minnesota, US

Robotics, Computer Vision, Reinforcement Learning

• Robotics, 3D Computer Vision, Reinforcement Learning

Bachelor

Tsinghua University, China

Mechanical Engineering

• Mechanical Design, Control Theory, Theortical/Fluid Mechanics

Diffusion-based Robot planning with Recursive Bayesian Filter [Title modified to ensure blind review]

Submitted to Robotics: Science and System (RSS) 2025.

Use diffusion model to predict human motion distribution. Plan robot path based on the prediction using a Bayesian Filter in a closed form.

Map-Aware Human Pose Prediction for Robot Follow-Ahead

2021 IEEE/RSJ International Conference on Intelligent Robots and Systems, (IROS 2024, Oral Presentation).

Address environmental-conditioned human pose prediction problem in the context of robot follow-ahead.

Learning to Play Pursuit-Evasion with Visibility Constraints

2021 IEEE/RSJ International Conference on Intelligent Robots and Systems, (IROS 2021).

Address pursuit-evasion problem in mainly two aspects: an multi-agent reinforcement learning; a representation design for belief state (contaminated region) to improve Markovian property.

A Novel Parallel and Self-adaptive Robot Hand with Triple-shaft Pulley-belt Mechanisms

International Conference on Intelligent Robotics and Applications (ICIRA) 2017

Design a noval parallel pinching and self-adapting grasping hand, TPM hand.

Reinfocement Learning for Autonomous Driving Internship

Plus.AI

Offline reinforcement learning for autonomous driving. Deployed trained agent into the ROS system with ONNX and TensorRT.

• Implement Offline Reinforcement Learning (Offline-RL) algorithms for the vehicle controller, including Decision Transformer, Conservative Q-Learning (CQL), etc. Benchmark with imitation learning results and classical controllers such as pure pursuit.
• Collect large-scale datasets for the RL agent training from the existing driving database. Decode data points of desired features from ROS bags, synchronize between topics, and pre-process the data.
• Integrate the trained agent into the vehicle ROS system in C++ by exporting the model to ONNX and loading it with TensorRT. Conduct multiple unit tests and evaluate the performance in the simulator.

Founder

SportsVision

Automatic highlight extraction for basketball games

• Used by 10K users; Deployed in 3 basketball courts with 8 cameras
• Install cameras around the basketball court and provide auto-clipped highlights for players.
• Automatic goal detection using Yolo for basketball and human detection and designed algorithm.
• Multiple object tracking (MOT) for assigning highlight clips to different users.
• WeChat mini-App as the front end for users to select and download the clips.

Closed-loop Diffusion Planning

[paper] Diffusion-based Robot planning with Recursive Bayesian Filter [Title modified to ensure blind review]

• Generative model for human motion prediction: Using diffusion models to predict the future human motion distribution, conditioned on the environment and past human trajectory.
• Closed-loop planning with posterior observation: Update the estimated human motion distribution using a recursive Bayesian filter and importance sampling technique. Plan the robot motion by updating the predicted distribution in a closed-loop manner with posterior human motion observations. Solve the robot planning problem.

Human Pose Prediction

[IROS 2024 Oral Presentation] Map-Aware Human Pose Prediction for Robot Follow-Ahead

• Human motion forecasting: Proposed a GRU-based method that predicts the human motion (3D skeleton poses) conditioned on the map information. Compared to the SOTA methods, we improve the accuracy while greatly shortening the inference time.
• Robot Follow-ahead: Build a robot that follows the human in the front. The robot can navigate through an indoor environment with an extbf{SLAM} module while predicting the 3D skeleton poses of a dynamic human.

View Planning for Dynamic Objects

Online view planning metric on dynamic object for high-fidelity 3D reconstruction

• View planning for 3D reconstruction: Propose viewing quality metric (Pixels-Per-Area) and design view planning strategy for tracking multiple dynamic actors online and 3D reconstructing offline.
• Drone system: Build a drone system that localizes a moving human, plans views to maintain viewing quality, and reconstructs offline by fusing consecutive views.
• Simulation: Evaluate the method in the Airsim (Unreal Engine) simulation environment and compare 3D reconstruction results with state-of-the-art methods.

Manpulation for Fruit-Picking

RL-based manipulation for object grasping and fruit-picking with tactile feedback.

• RL-based manipulation: Train a Reinforcement Learning (RL) agent for a grasping task (fruit-picking) with tactile feedback. Benchmark with RRTs (implement from scratch) and other motion planning algorithms from the Open Motion Planning Library (OMPL). Simulate the manipulation in V-REP (CoppeliaSim).
• Hardware integration: Integrated tactile sensors into the manipulation system and conduct hand-eye calibration for the on-hand camera.

3M Graduate Student Fellowship (4-year)

University of Minnesota

School Year Academic Scholarship (top 10%)

Tsinghua University

Civil Engagement and Sports Scholarship

Tsinghua University

President Elected

Study and Career Development Center in Student Union

Summer camps for three consecutive years to empower 100+ “left-behind” Chinese children whose parents are absent due to economic displacement, with a team of 40+ volunteers. Raised Funds.

• Led a committee of 20 members to improve departmental career services for 400+ students.
• Launched a career development mentor program which connects 20 pairs of students and alumni.
• Established an alumni database which includes career information for 300+ alumni.

Founder

'Cuckoo Project' Annual Summer Charitable Program

Summer camps for three consecutive years to empower 100+ “left-behind” Chinese children whose parents are absent due to economic displacement, with a team of 40+ volunteers. Raised Funds.

• Organized summer camp in 3 consecutive years to empower 100+ “left-behind” Chinese children whose parents are absent due to economic displacement, with a team of 40+ volunteers.
• Raised 150,000 RMB (approximately $22,000) and set up logistics for the summer program by building relationships with partners from government, national non-profit and private sector.
• Developed and managed evaluation procedures which culminated into annual project reports.