Left-invariant EKF for Drone Localization

in-EKF, Localization, Sensor Fusion

An invariant EKF is implemented for a quadruped in Matlab and C++. We used self-generated fake data and Zurich Urban Aerial Vehicle Dataset to verify our filter. For more details, please refer to the report in our github page.

Close Project

Learning Based Sampling for RRT*

RRT*, Conditional Variational Autoencoder (CVAE), Machine Learning

Compared with traditional uniformly sampled path planning approaches, a biased sampling method which learns from demonstrations and environmental restrictions would be preferable due to its efficiency and generality, as it can capture the obstacles and some sense of dynamics in the system. In this project, we have combined RRT∗ algorithm and a conditional variational autoencoder (CVAE) trained to learn from the latent dimensions and propose samples accordingly. We have generated and trained on our own datasets, including 4 maps and 26985 sampling states. The sampling states are generated with hybrid A*.

Assuming latent variables are Gaussian, the CVAE is trained to maximize the conditional log likelihood. The network contains a three-layer encoder and a three-layer decoder. The input to the encoder includes the grid obstacle map, successful planning states, intial and end point of the trajectory. In the offline test phase, only the decoder is utilized.

The modified planning algorithms are then applied on the vehicle parking path planning, and compared with the uniformly sampled RRT∗ and bi-RRT∗. A substantial boost is observed in path planning performance with this inclusion of biased sampling network. For RRT∗, the number of nodes extended, elapsed time and path length have dropped by 72.5%, 91.4% and 18.0% respectively. For bi-RRT∗, they have dropped by 55.1%, 70.4% and 17.7% respectively.

Close Project

Mobile Robot Maze Exploration

Robot Platform Development, Forward/Inverse Kinematics, Particle Filter, Path Planning

The SLAM (with particle filter), path planning, exploration and block retrieval functions are designed and implemented on a mobile robot. The robot is equipped with BeagleBone Green for motion control and Raspberry Pi for SLAM and arm module.

With these subsystems, the robot is capable of exploring in unknown environment, constructing a map of the environment and retrieving blocks in it. The occupancy grid mapper in the SLAM system is able to generate grid maps at the resolution of 0.05 m, the Monte Carlo localization algorithm is able to reach a mean accuracy of 0.05 m in translation and 0.08 rad in rotation.

This system completed four tasks which included: 1) completing four circuits around a convex arena, 2) detection and relocation of six blocks within convex area, 3) exploration and mapping of an unknown environment, and 4) detection and relocation of 8 blocks scattered across a map.

Close Project

Drainage Inspection Robot

Robot Platform Development, Object Detection, 3D Reconstruction

We develop an autonomous drains inspection robot platform that detects water and debris by using its onboard vision system. I mainly worked on debris detection algorithm utilizing 3D point cloud constructed from the stereo system and deployment of the visual module on the robot.

The debris detection algorithm is separated into three different parts. Firstly, the point cloud of the scene is reconstructed based on the stereo vision cameras’ output. Then, two sections of points along the drain are extracted and projected onto the X-Y plane. Secondly, an iterative constrained line searching algorithm is applied to obtain the drain ground markers, which represent the ground’s shape. Subsequently, this shape is extended to 3D dimensions to remove the cloud’s points along the ground planes. Finally, the rest of the points are clustered and classified as debris or obstacles based on their positions to the ground markers.

The robot was tested in a 100-meter drain in Singapore. 99% and 91% of the water and debris were correctly detected, respectively. Furthermore, an obstacle obstructing the robot’s progression inside the drain was recognized 90% out of 20 trials.

Close Project

Quadruped Nonlinear Trajectory Planning

Simulation, Motion Planning, Nonlinear Programming

A offline nonlinear trajectory planning approach utilizing TOWR for our quadruped Pholus was implemented. The trajectory for four end-effectors are calculated and simulated with Gazebo. We are using single rigid body dynamics for the dynamic model and a cartesian space kinematic model bouding the motion of the end effectors.

Close Project

PointNet for Point Cloud Classification

point cloud, classification, supervised learning

We reimplemented PointNet for object classification with our own modifications. The results are evaluated utilizing ModeNet40.

Modifications involve incorporating spatial transform networks to align the data, adding normal channels and including farthest point sample layer for better representation of geometric data.

Close Project