Multi-Modal Trajectory Prediction & Imitation Learning

Obstacle trajectory prediction is crucial for modern autonomous vehicles. In this project, we developed multi-modal trajectory prediction. That is, the deep learning model can predict multiple possible trajectories and their corresponding pobabilities, considering map, topology and surrounding agents’ motions. Those information are very useful for downstream control & planning module.

I also also design an archor-based imitation learning algorithm which learns from drivers’ demonstrations and can automatically generate multiple possible trajectories with diffusion model, given anchor points and driver’s historic motions. The planned trajectory can be tracked by downstream control module. In practice, we can sample future points on centerlines as anchor points.

Continuous-time Continuous-space MAPF

In this project, I developed a large-scale continuous-time continuous-space MAPF system. Based on SIPP, priorized planning, and large neighborhood search, the algorithm is very flexible, efficient and has the anytime behavior.

Lifelong MAPF

In lifelong MAPF setting, each agent will receive a new goal upon the completion of the current goal. Based on CBS, RHCR and prioritized planning, my MAPF system can keep generating new paths for agents while maintaining safe and handling the direction of agents in different maps in a timely manner.

Hierachical mapping-based Object Navigation

In this robot system, the mobile robot can detect objects, explore the building and navigate to the requested objects once they are detected. A multi-layer hierachical map is built online, which contains the object layer, occupancy and the room layer. The hierachical map provides critical information for frontier-based exploration, path planning, and trajectory tracking. In this project, we also use LLM to guide robot’s navigation between rooms.

Leviathan’s Treasure

It is an online multi-player asymmetric action game, powered by Unity and proto. One player plays the role of monster, and three players playes the role of hunters which include knight, mage and soldier. Each character has different abilities. For example, the monster can flying, teleport and throw fireballs. The goal of hunters is to collect monter’s treasure while the monster is trying to eleminate all hunters.

Topics:

MuLTIPLAYER

aCTION

fANTASY

asymmetric

Security robot Temi

Powered by deep learning and Android programming, this security robot is designed to monitor laboratories with minimal human intervention. It can communicate directly with users to securely collect their IDs, passwords, and voice data for registration and login. Additionally, the robot can detect and localize abnormal sounds, navigate autonomously to the corresponding location, and record video footage to issue timely alerts.

Topics:

Chatbot

Robot

Sound localization

Lidar-based SLAM Game

In this project, we implement a LiDAR-based ICP algorithm to build the map in real-time. The constructed map is then transmitted to a remote computer for visualization. Additionally, we developed an app capable of sending commands to the robot and displaying critical sensor information, such as ultrasonic readings. We also designed the PCB and other electronic components essential for the mobile robot’s operation. As a result, users can remotely guide the robot through unfamiliar mazes—a truly impressive capability.

Some interesting robotics work