Mapping and Map Representations [ Mapping slides ] :

Covering mapping and map representations, I presented essential concepts: 2D occupancy grids for environment breakdown, 3D occupancy maps for 3D spaces, depth maps for depth info, point clouds for 3D data, height maps for elevation, landmark maps for distinct features, distance maps for path planning, surface representations for geometry, and multi-map representations for versatile task optimization in robotics and computer vision.

Simultaneous Localization And Mapping [ SLAM slides ]:

I explained SLAM comprehensively, starting with online methods like EKF SLAM and Fast SLAM, moving on to pose graph optimization and loop closure. They then discussed full SLAM techniques like Graph SLAM, ORB SLAM, Factor Graphs, and LIO-SAM, highlighting their role in mapping and navigation. The session culminated with a practical Amazon Astro robot demonstration, illustrating the real-world application of SLAM in autonomous systems

A-star Planner: 8-connected nearest neighbors based a-star planner finds the shortest path from the start to the goal. Here, a Pr2 mobile manipulator is shown to find the shortest path and traverse along that path even with challenges.

RRT Planner: Rapidly-Exploring Random Trees is a motion planning algorithm that efficiently explores the configuration space of a robotic arm to find a feasible path between start and goal configuration. Here, a PR2 mobile manipulator is used to find the feasible path and also find the shortest path using the trajectory obtained using RRT-connect.

Point Cloud Registration: Here, Point Feature Histogram which is a descriptor used in point cloud analysis is utilised to characterize the local surface properties of the point cloud. This can be used in correspondence estimation between the source and the target point clouds and align it using Iterative Closest Point algorithm.

This project leverages an RGB-Lidar camera system to detect blocks and estimate depth, incorporating Forward Kinematics (FK), Inverse Kinematics (IK), and motion planning algorithms to enable autonomous grasping using a 5 Degree of Freedom (DOF) robotic arm. The RGB component captures object color and texture, while Lidar measures distances, facilitating 3D block detection. Depth estimation creates a 3D map for precise positioning. FK computes the arm's end-effector position, while IK calculates joint angles for grasping. Motion planning generates collision-free paths, culminating in autonomous block grasping, representing a significant step in robotics for complex tasks in unstructured environments.

In this project, We developed an M-Bot equipped with wheel encoders and an IMU for odometry estimation, and then integrated 2D lidar sensor data with a particle filter localization system, alongside A* path planning and frontier exploration techniques to achieve Simultaneous Localization and Mapping (SLAM). The wheel encoders and IMU provide essential data for tracking the robot's movement and orientation, while the 2D lidar sensor captures environmental information. The particle filter localization method utilizes this sensor data to estimate the robot's precise position and orientation in real-time. Furthermore, A* path planning is employed to generate optimal paths for the robot to navigate its surroundings, ensuring efficient movement. Additionally, frontier exploration techniques are utilized to autonomously identify and explore unknown areas, contributing to comprehensive SLAM capabilities that are crucial for autonomous mobile robotics in diverse environments.

Simultaneous Localization and Mapping (SLAM) is not robust in dynamic environments, where objects move and cause changes in the surroundings, its effectiveness is limited. It is particularly challenging to use SAM methods that depend on visual or point cloud data, as the dynamic objects can be registered in multiple scans. To address this issue, we propose a method that combines Sparsely Embedded Convolutional Detection (SECOND) and LiDAR Inertial Odometry Smoothing And Mapping. Our approach involves masking out the point cloud features detected by SECOND in the Velodyne scan. Then we occlude the features and update it to LIO-SAM to help it develop the map and state estimation. Our experimental results demonstrate the effectiveness of our approach, which achieves good? accuracy in localization compared to the original LIO-SAM method. By enabling SLAM to work more effectively in dynamic environments, this could enhance the safety and reliability of various autonomous systems.

The set of objects to be stacked is fixed and is composed of the red, green and blue rectangular polygonal shapes and the cyan box shape. Since the objects are set very close, they need to be pushed to a certain distance before grasping them. The pose of all the objects are given and the code for pushing and grasping the box was done and simulated as shown.

In this project, the primary challenge was to perform 3D object detection using a single 2D image, a critical task for ensuring safe navigation in autonomous driving systems. Inspired by Mousavian et al., the approach involved switching the 2D detector from Faster-RCNN to YOLO and changing the regressor model from VGG to MobileNetV2. The core of the project was the creation of 3D bounding box estimation, allowing the system to better understand its surroundings. By utilizing YOLO for 3D detection and MobileNetV2 for the regressor model, we achieved improved performance and capabilities. The effectiveness of the methods was quantified by constructing a bird's eye view of the images and evaluating the intersection of union.