Dhruvil Katharotiya

Projects

Research and engineering work from my Master's at THI, plus personal projects.

Task-aware Grasp Selection & Path Planning (Thesis)
Evasive Maneuvers for Automated Vehicles
Q-Learning in a Custom GridWorld
Eco-Driving at Signalized Intersections
Flash Card Trainer Web App

Task-aware Grasp Selection and Path Planning for KUKA LBR iiwa Cobots

Master's Thesis · Technische Hochschule Ingolstadt · ongoing

Robot arm grasping an object and following a planned path

A robot should grasp an object in a way that fits what it has to do next. A mug picked up to be poured needs a different grasp than one picked up to be stacked.

My thesis studies how to choose grasps with the downstream task in mind and how to plan collision-free motions for the KUKA LBR iiwa 7-DoF collaborative arm.

Topics: robot manipulation · grasp planning · motion planning · KUKA LBR iiwa

Local Planning and Evaluation of Evasive Maneuvers for Automated Vehicles

Master's Project · THI · 2024 · team of four

Lane-change trajectory around a slower vehicle

A local planner makes short-term, real-time decisions so an automated vehicle avoids collisions while staying smooth and following traffic rules. We built a Python simulation that decides when the ego vehicle should change lanes and generates a smooth trajectory for doing it.

  • Scenario-based collision detection and lane-change logic from relative positions, velocities and safety distances, validated across 50+ simulated traffic scenarios.
  • Smooth lane-change trajectories with cubic-spline interpolation; smoothness was the main challenge.
  • Polynomial Regression and Decision Tree models to predict vehicle positions and explain maneuver decisions, with sensitivity analysis across 30–130 km/h.

Tools: Python · NumPy · SciPy · scikit-learn · Matplotlib

Q-Learning Agent in a Custom Gymnasium GridWorld

Course Project · THI

Heatmaps of learned Q-values for the actions Up, Down, Right and Left

I built a 6×6 grid environment from scratch on the Gymnasium API, with obstacle ("hell") states, a black hole and a goal, rendered with Pygame. A tabular Q-learning agent learns to reach the goal while avoiding hazards.

  • Custom gym.Env with discrete actions, reward shaping and episode termination.
  • ε-greedy exploration with decay (ε from 1.0 to 0.1), discount γ = 0.99, trained for 20,000 episodes.
  • Visualised the learned Q-table as per-action heatmaps (left) to check that the policy steers clear of hazards.

Tools: Python · Gymnasium · Pygame · NumPy · Matplotlib

Eco-Driving at Signalized Intersections

Scientific Seminar · THI · 2024 · supervised by Anna Lena Schlamp

Speed profile approaching a traffic light: stop-and-go versus eco-driving

A literature review of how vehicles can approach traffic lights more efficiently by anticipating signal changes, keeping speed steady and avoiding unnecessary stops, to save fuel and cut emissions.

  • Compared Dynamic Programming, Model Predictive Control, Gaussian Process Regression and Soft Actor-Critic reinforcement learning.
  • Examined integrated evaluation setups that couple traffic, V2X network and driving simulators, and predictive cruise control (Eco-PPCC).
  • Discussed real-world limits, especially without vehicle-to-infrastructure (V2I) communication.

Topics: reinforcement learning (SAC) · Gaussian processes · MPC · connected vehicles

Flash Card Trainer Web App

Personal Project

Stack of flash cards with a review schedule

An adaptive learning web app with spaced repetition: cards you struggle with come back sooner.

  • Flask backend with user accounts, progress tracking and spaced-repetition scheduling.
  • Containerised with Docker; automated testing and deployment through GitLab CI/CD.

Tools: Python · Flask · Docker · GitLab CI/CD