Projects

Project Overview:

Objective: Aimed to develop and refine Kalman and Particle filters for accurate localization in autonomous robotics.

Development and Testing: Developed and implemented Kalman and Particle filters, creating scenarios in PyBullet for testing. Conducted comprehensive integration and testing with robot motion models in varied conditions.

Result: The comparative analysis of the Kalman and Particle filters revealed crucial insights. The Particle Filter demonstrated resilience in non-linear environments, emerging as a versatile choice for state estimation in complex scenarios.

Tools Used:

Project Overview:

Odometry System Refinement: Calibrated PID controllers to enhance the robot's odometry-based positioning and mapping, utilizing wheel encoders and IMUs for improved accuracy.

Particle Filter-Based SLAM: Developed a SLAM system with AMC/Particle-filter on the Mbot platform for dynamic localization, supported by a 2D Occupancy-Grid mapping method.

Motion Control Optimization: Advanced motion control using PID algorithms, leveraging odometry data and IMU sensors for precise maneuvering and trajectory maintenance.

Exploration Algorithm: Deployed a frontier-based algorithm for exploration, using 2D RP Lidar to detect reachable frontiers, and implemented the A* algorithm for optimal path planning.

Computer Vision Integration: Employed an RGB camera for environmental perception, facilitating cube orientation detection with Apriltag recognition for interaction tasks.

Robotic Manipulation: Custom-designed a gripper for warehouse-style cube manipulation, integrating it with the robot's autonomous systems for task execution.

System Integration and Navigation: Combined odometry, SLAM, vision systems, and a custom gripper with a Jetson Nano, enhancing the two-wheeled differential drive robot’s autonomous navigation and operational efficiency.

Tools Used:

Project Overview:

Advanced Object Detection Technique: Implemented object detection in cluttered environments using Region Proposal Networks (RPN) and Mask R-CNN.

PROPS Dataset Utilization: Employed the PROPS dataset, comprising annotated bounding boxes for 10 object classes, enhancing detection accuracy.

Optimization Techniques: Implemented Non-Maximum Suppression (NMS) for refining detection by selecting high-scoring boxes and eliminating lower-scored ones.

Loss Calculation: Computed box regression loss and objectiveness loss, crucial for model accuracy.

Tools Used:

Project Overview:

Dynamic Model Implementation: Developed a robust dynamic model for robot planning and control, enabling precise object manipulation within a PyBullet simulation environment.

SE2PoseLoss Prediction: Innovated a pose prediction method with SE2PoseLoss, ensuring accurate forecasting of object placement post-robotic action.

Residual Dynamics Learning: Advanced the learning of system dynamics through Residual Dynamics Learning, streamlining the complexity of mapping predictions for improved learning efficiency.

MPPI Controller Modeling: Modeled an MPPI controller to strategically plan a sequence of actions, optimizing the robot arm's path to the goal configuration.

Neural Network for Dynamics: Trained a three-layer neural network with ReLU activations to model the pushing dynamics, enhancing the robot's interactive performance.

Task-Specific Requirement Encoding: Defined a cost function that encoded task-specific requirements, enabling the robot to minimize goal distance and avoid obstacles efficiently.

Tools Used: Python, PyBullet, NumPy, Eigen.