LLM meets robotics (industry project)
Supervisor: Benedikt Kreis

The goal of the project is to pick unknown objects from a moving conveyor belt and sort them into different bins. Instead of hard coded control commands, the robot has to exploit capabilities of LLMs (Large Language Models) and VLMs (Visial Language Models). This project is offered in collaboration with a local industry partner.

Robotic sorting using Reinforcement Learning
Supervisor: Ahmed Shokry

The task is to sort cubes according to some features using a robotic arm. The robot will use Reinforcement Learning and information from RGB-D camera to identify, locate, and pick the cubes and place them in the correponding bins.

Learning-based navigation among dynamic obstacles
Supervisor: Jorge de Heuvel 

Using state-of-the-art deep reinforcement learning methods, you will teach a robot to navigate among static and dynamic obstacles. Here, a neural network is trained that directly controls the robot based on observations through a LIDAR sensor. You will design your own learning architecture that includes the network design for a solid understanding of the for the robot partially observable environment, and the reward function for informative and efficient learning of the navigation objectives. After the robot successfully has trained in Pybullet-based simulation, the goal is to transfer the policy onto a real robot using ROS.

Quadruped for obstacle courses
Supervisors: Murad Dawood/Shahram Khorshidi

The goal of this project is to enable the quadruped to navigate through obstacle courses, while using perception to adapt to different velocities, obstacles, and staircases.

Autonomous racing
Supervisor: Nils Dengler

The task is to learn a reinforcement learning agent to autonomously drive through a given race track. For simulation training we use the pybullet simulation given in [1]. After successfully training the agent, we deploy it on a real F1/tenth cart and drive a track in the hallway of our building.

[1] https://github.com/gonultasbu/racecar_gym-1

Rock, paper, scissors
Supervisor: Nils Dengler

The task is to implement a version of Rock Paper Scissors on our 5-finger Psyonic hand [1].

[1] https://www.psyonic.io/ability-hand

P2G: Predict to Perceive and Grasp for Mobile Manipulation
Supervisor: Rohit Menon

This project aims to use shape prediction for enabling a mobile manipulator to aid both next best view planning as well as for grasp pose detection. With shape prediction, the robot should be able to complete the manipulation task with higher success rate and lower time.

Find the viewpoint!
Supervisor: Sicong Pan

The task is to find the viewpoint of the given RGB image within the context of an eye-in-hand tabletop configuration.

Social force model for crowd simulation
Supervisor: Subham Agrawal

The task is to implement social force model (SFM) for crowd simulation and visualize it in a 3D environment.

Robot navigation using LLM
Supervisor: Xuying Huang

This project aims to use a large language model (LLM) to achieve robot navigation. The robot should be able to understand spoken instructions, process the command using the LLM, and convert it into specific actions (such as moving to the kitchen).