Ekansh Gupta

Differentiable LOAM

Problem: Developing a differentiable variant of LOAM (Lidar Odometry and Mapping) for end-to-end learning in autonomous vehicles.

Solution:

• Developed a differentiable version of LOAM, enabling sensitivity analysis of point cloud perturbations on map generation and path planning.
• Refined differentiable point cloud sampling algorithms.
• Implemented soft selection for scan registration and point selection from LiDAR data.
• Designed and trained an attention-based neural network to assign weights to each point based on their relevance for ICP.
• Collected pointcloud data using the CARLA simulator for training the network.

Impact:

• Enhanced autonomous navigation capabilities in environments with controlled dynamic obstacles.
• Improved overall mapping efficiency and robustness of LOAM in real-world scenarios.
• Integrated and tested the system on MathWorks' autonomous vehicle platform, demonstrating real-world self-driving capabilities.
• Attention-based weighting helped mitigate the impact of dynamic obstacles, significantly improving mapping efficiency.

Technologies: Python, PyTorch, ROS, CARLA Simulator, Deep Learning, perception, point cloud processing.

Differentiable LOAM implementation for ADAS

Unmanned Aerial Vehicle

Problem: Design an autonomous system for a passenger drone capable of safe, efficient navigation in complex urban environments with dynamic obstacles.

Solution:

• Developed and implemented B-spline based path planning algorithms for smooth and efficient drone navigation.
• Integrated various sensors for enhanced environmental awareness and obstacle detection.
• Utilized the local support property of B-splines to enable real-time path modification in response to dynamic obstacles.

Impact:

• Achieved collision-free navigation in simulated urban environments with dynamic obstacles.
• Enhanced the overall safety and efficiency of passenger drone operations.
• Successfully contributed to a $1 Mn, government-funded project with significant implications for urban air mobility.

Technologies: ROS, Gazebo, Python, Sensor fusion, SLAM, Pixhawk flight controller, Bsplines.

Test flights at IIIT Hyderabad and IIT Hyderabad. As seen, after going to offboard mode, the drone detects an obstacle in Rviz and changes track from the red one by manipulating the control points of the Bspline curve - generating a smooth trajectory.

Autonomous Wheelchair

Problem: Develop an autonomous wheelchair system capable of navigating complex indoor environments like hospitals and airports using visual-language commands.

Solution:

• Spearheaded the hardware implementation of an autonomous wheelchair prototype.
• Implemented AMCL (Adaptive Monte Carlo Localization) for precise position estimation and RTABmap (Real-Time Appearance-Based Mapping) for robust 3D mapping of indoor environments.
• Helped in training a Vision-and-Language Navigation (VLN) model, enabling natural language inputs for intuitive control.
• Optimized the system for accurate navigation in challenging indoor settings with dynamic obstacles and varying layouts.

Impact:

• Achieved high-precision indoor navigation, enhancing safety and reliability in complex environments.
• Enabled intuitive control through vision language commands, improving accessibility for users.

Technologies: Python, ROS, VLM based Deep Learning, AMCL, RTABMaps, SLAM, sensors, actuators, and Control Systems.

Video demonstration of the project and its tracking on rviz. In this, we tested the bot's motion through congested door so had to give each waypoint separately.

Autonomous Quadruped

Problem: Develop an advanced quadruped robot with Level 4 autonomy, focusing on Reinforcement Learning for gait pattern mastery and complex physics understanding.

Solution:

• Led the design and implementation of an indigenous quadruped robot funded by BITSAA with a grant of $30,000.
• Developed a Reinforcement Learning (RL) environment for training the quadruped. Simulated and trained the RL model to autonomously learn complex physics and master diverse gait patterns.
• Programmed and optimized various gait patterns using DDPG.
• Integrated deep RL agent (DDPG) with ROS architecture to improve stability and adaptability.

Impact:

• Created a long-term learning platform for students to apply theoretical concepts in robotics.
• Pioneered an RL-focused approach, differentiating from kinematics-heavy designs like Boston Dynamics' Spot.

Technologies: Python, OpenAI gym, PyTorch, ROS, Reinforcement Learning, Pybullet.

Policy trained in DDPG using Bezier curves as the base policy.

Path planning for 7 dof panda arm

Problem: Develop sophisticated control and path planning algorithms for a 7-DOF Panda robot arm to perform precise movements and obstacle avoidance.

Solution:

• Implemented Cartesian control and inverse kinematics using numerical methods for precise arm manipulation.
• Developed RRT (Rapidly-exploring Random Tree) path planning algorithm for obstacle avoidance and goal-reaching tasks.
• Utilized the null space of the Jacobian matrix for end effector stabilization while manipulating the extra degree of freedom.

Technologies: ROS2, RviZ, Forward and Inverse Kinematics, Path planning.

Tango - Parallel Linked Quadruped

Problem: Developing a fast legged robot.

Solution:

• Modeled, 3D printed, assembled, and wired up a parallel linkage quadruped for real-world validation.
• Deployed a Reinforcement Learning (RL) model for adaptive gait generation using Isaac Sim and ROS2 architecture for teleop control.
• Implemented strategies to close the Sim2Real gap, ensuring consistent performance between simulation and reality.

Impact:

• Due to the modern design, the bot achieved a walking speed of over 18 cm/s.
• The parallel leg linkage helped in faster joint movements and increased the robustness of the bot due to stable COM.

Technologies: ROS2, Pybullet, Sim2Real, Reinforcement Learning, Isaac Sim, 3D printing and CAD.

Chatbot for Crustdata API

Problem: Develop an intelligent chatbot capable of providing accurate responses based on Notion documentation and Slack channel history.

Solution:

• Implemented a specialized web chatbot using Langchain with RAG, integrating Notion documention and examples along with additional knowledge from Slack.
• Deployed the chatbot on Slack for instant query resolution.

Technologies: Langchain, RAG, NLP, APIs, LLM.

Flipkart Grid 2.0

Problem: Develop an autonomous hexacopter capable of precise indoor navigation for the Flipkart Grid challenge.

Solution:

• Implemented a computer vision system using Realsense camera and OpenCV to accurately calculate the center coordinate position within the camera frame.
• Designed and implemented a PID control system for precise movement towards specified goals.

Impact:

• Achieved top 5 ranking nationwide among 120+ registered teams in the Flipkart Grid challenge.
• Demonstrated advanced capabilities in autonomous drone navigation and precision control.

Technologies: Python, ROS, Gazebo, OpenCV, Control Systems.

Extended Kalman Filter

Implemented Extended Kalman filter for accurate 2D state estimation using noisy sensor data in ROS2.

Solar panel cleaning bot

Under Dr. Manoj Soni's guidance, I led the development of an adaptive solar panel cleaning robot, securing Rs 40,500 in funding for prototyping. Our approach evolved from an initial X-Y core mechanism designed in MATLAB Simulink to a more advanced, efficient design.

SAP Labs

IIIT Hyderabad

Campus Binge

Sirius Motor sports

Columbia University, New York

Bits Pilani Pilani Campus