1 of 45

Project topics in artificial intelligence and robotics

real.itu.dk - 2025

2 of 45

Embodied affective

Robotics

Robotics

Andrés Faiña, CS

anfv@itu.dk

Kasper Støy, CS

ksty@itu.dk

Mechatronics

Bioinspired and swarm robotics

Payam Zahadat, CS

paza@itu.dk

Machine

Learning

Lab Automation

Neural Networks

Sebastian Risi, DD

sebr@itu.dk

People

Nicolas Bessone,

CS nbes@itu.dk

Jonas H Jensen, CS

jhaj@itu.dk

Djordje Grbic, DD

djgr@itu.dk

Joachim Winther Pedersen, DD jwin@itu.dk

Elias Najarro, DD

enaj@itu.dk

Rodrigo Moreno Garcia, CS

rodr@itu.dk

Morten Roed

Frederiksen

mrof@itu.dk

3 of 45

Four thesis elements

Build robot Prototype

Simulation

Control software

Artificial intelligence

Application domain

Experiments

Existing

Robot

Research

topic

4 of 45

Overview of topics

  • Modular robots and robot design
  • Swarm robots
  • Social Robots & Embodied Affective robots
  • Emergence, Open-endedness, and generative models
  • Games, multi-agent and bio-inspired optimization

5 of 45

Bio-inspired and swarm robotics

learn from nature’s amazing solutions

Take inspiration from natural systems

Design algorithms for robotic systems

Contact: Payam - paza@itu.dk

6 of 45

Evolutionary Robotics:

Evolving a General Communication Space for Robotic Swarms

a novel, general approach to evolve reliable and explainable communication in swarms of robots, and in particular the interplay between behavior and communication.

Contact: Payam - paza@itu.dk

7 of 45

Dynamic collectives: static formations consisting of internal dynamics

Contact: Payam - paza@itu.dk

Develop algorithms for swarms of autonomous robots.

The overall shape of the collective is static, but those who form it are dynamic (circulating within the swarm or constantly leaving and joining).

8 of 45

Maritime Robotics

Applying swarm and modular robotics to develop new approaches in container shipping.

Contact: Payam - paza@itu.dk

9 of 45

Robots vs Anxiety

  • Enable pocket-sized robots to make it’s users challenge themselves to improve?
  • Enable robots to communicate such challenges?
  • Goal oriented LLM

Contact: Morten (mrof@itu.dk)

10 of 45

Can demands and caretaking needs spark a stronger affective connection with therapeutic robots?

This project will investigate if introducing tamagotchi-inspired demands to a pocked-sized therapeutic robot can catalyze stronger motivation to interact with the robot.

Contact: Morten (mrof@itu.dk)

11 of 45

Tactile robot communication

  • Can robot to robot communication be a viable alternative to modern chat platforms?
  • What can it offer that these can’t?
  • Can physical and tactile communication improve the connectedness of a group? (virtual tactile handshakes?)

12 of 45

Modular robots

Contact: Andres Faina (anfv@itu.dk)

Office: 2F18

Robotic Modules

Modular robots built by assembling modules

  • Potential projects
    • Add sensors (real or simulated)
    • Add different types of modules
    • Improve modules or connection mechanism
  • EMERGE modular robot: link

13 of 45

Evolutionary modular robots

Contact: Andres Faina (anfv@itu.dk)

  • Improve MAP-ELITES algorithm for modular robots
  • Evolve different robotic morphologies with different controllers
  • Test them in a automated platform (using a robot arm if needed)
  • Compare real results with simulation results
  • Final goal: Obtain evolved robots able to work in reality!

14 of 45

Evolving modular robots in GPU

  • Robot morphologies have been evolved in simulation but mostly using CPUs.
  • Can you learn controllers for different morphologies using GPUs to speed up the training? Is it faster than CPU with a large number of collisions/robots?
  • Use MuJoCo XLA (MJX) (Robotics Simulator with GPU support)

Contact: Andres Faina (anfv@itu.dk) and Rodrigo Moreno (rodr@itu.dk )

https://vimeo.com/540894612

15 of 45

Optimizing modular robots in the real world with bayesian methods

  • Robot controllers can be obtained in simulation but they behave differently in the real world
  • Use Bayesian methods to build a surrogate model from simulated data and use it to optimize a robot controller in few real world tests.
  • Use scikit-optimize python library

Contact: Andres Faina (anfv@itu.dk) and Rodrigo Moreno (rodr@itu.dk )

https://vimeo.com/540894612

16 of 45

Robots that are able to grow

  • The development of the body helps humans to learn
  • Can we explore the advantages of morphological development in robots?
  • Approaches:
    • Build a robot that is able to grow and replicate tests performed in simulation
    • Use modular robots to test morphological development in robotics (Grow = Add more modules to the robot) (Simulation and/or physical tests, modules already built!)

Contact: Andres Faina (anfv@itu.dk)

17 of 45

Active inspections in factories

  • Projects performed in collaboration with Novo Nordisk and Helix Lab
  • Potential fellowships available
  • Projects:
    • Manipulation tasks with a Spot robot (e.g., tidy-up of the floor´s factory or assembly tasks)
    • Use machine learning to detect if surfaces are clean or if valves are open or closed
    • Drone positioning systems using using UWB modules and LIDARs

Contact: Andres Faina (anfv@itu.dk)

18 of 45

Valve Identification and Grasping for a Quadruped Robot in an Industrial Setting

  • Identify various types of valves in diverse industrial environments, including varied lighting and distance conditions.
  • Design a reliable and adaptable grasping method that can handle different valve sizes and shapes while ensuring the robot’s stability during operation.
  • Ensure that the robot can operate safely in proximity to human workers and sensitive machinery without causing disruptions or accidents.
  • Complex Environments: Manage obstacles and dynamic elements in industrial settings that may affect the robot's navigation and operation.

19 of 45

Design and Simulation of a Quadruped-Based Deployment Strategy for an Aerial Inspection Drone

  • Design a mechanism to carry a small UAV on top of a quadruped robot while maintaining an appropriate quadruped performance.
  • Develop a mechanism for the deployment of the UAV from the quadruped robot's end effector.
  • Developing a control modelling and manipulator path planning algorithm to deploy the drone in a specific pose.
  • Developing a control algorithm to grab the drone mid-air with the manipulator.
  • Research in digital twins.
  • Testing the solution in simulation and verifying in real life.

20 of 45

UAV System for Active Inspection in the Pharmaceutical Industry

  • Improve the UAV prototype capable of flying in GPS-denied indoor environments from a previous ITU student.
  • Reduce the weight and optimize the structural design.
  • Port the flight controller from iNav to PX4 or Ardupilot and integrate with ROS 2.
  • Add simulation support in Gazebo.
  • Improve the control algorithms to fly close to walls.
  • Design a protective sphere for 3-axis navigation.

21 of 45

Automated design of machines

  • Designing machines in CAD is an iterative process and time consuming
  • Can we automate the process using AI?
  • Approaches:
    • Identification of mechanical components, DOF and motion in existing designs (using ANNs)
    • Building devices from high level specifications and a library of components (APIs available for Fusion 360 and FreeCAD)
    • Linter for mechanical designs

Contact: Andres Faina (anfv@itu.dk)

22 of 45

Educational electronic simulator

  • Electronic is complex to teach and simulators do not provide useful guidance to students. Can we use simulators (e.g., Fritzing link) to provide better feedback to the students?

Contact: Andres Faina (anfv@itu.dk)

DO NOT TRY THIS CIRCUIT!

  • Some ideas to implement:
    • Interactive feedback (explain errors)
    • Automatic correction of MAs
    • Measure how they help learning

23 of 45

Evolved Soft Robots

A lot of soft robots have been evolved in simulation, but there is not a fast way to build them

Can we make modular soft robots to build the evolved robots?

Can we create robots that are able to grow?

Contact: Andres Faina (anfv@itu.dk)

24 of 45

Swarm robotics

  • Design the lowest cost robotic swarm possible.
    • Cost <10$
    • Easy to charge:
      • Self-recharging
      • Solar cells
    • Design to be manufactured without manual assembly

Contact: Andres Faina (anfv@itu.dk)

25 of 45

Origami Robotics

Explore the creation of 3D printed origami mechanisms for use in soft robotics.��Design new origami mechanisms

Investigate print in place solutions for robotic end effectors

Contact: Bailey Dacre baid@itu.dk

26 of 45

Morphing buoyancy robot

Liquid

Contact: Rodrigo Moreno rodr@itu.dk

Applications:

Chemistry, measure features in liquids

Underwater monitors/cameras

How can a robot change its whole shape to change its buoyancy?

27 of 45

Vibration sensing for liquid level in vessels

  • How can liquid levels be identified from listening to a vibrating container with machine learning?

Similar techniques for identifying solids and changes in solids

Contact: Rodrigo Moreno rodr@itu.dk

28 of 45

Smart, portable testing of water treatments with automatic modules

  • In collaboration with Alumichem and the Helix lab
  • Potential fellowships available

Contact: Rodrigo Moreno rodr@itu.dk

29 of 45

From robot to API in hardware

How to automate the generation of controller Arduino libraries from robot specifications:

Datasheet of the device + requirements ->Auto generated Arduino code -> Verification

Some past work in drivers for operating systems:

Contact: Rodrigo Moreno rodr@itu.dk

30 of 45

Chemistry with Modular Reconfigurable Lab Automation

Dive into control, simulation, mechatronics, and machine learning with our system.

—no chemistry background needed.

Jonas Jensen jhaj@itu.dk

31 of 45

Large-scale co-evolution in a Massively Multi-agent Game environment

  • Use the Neural MMO environment (https://openai.com/research/neural-mmo) to co-evolve populations of thousands/millions of neural network-controlled agents.
  • Under what circumstances would interesting group dynamics evolve? What environmental factors drive agents to form different species or when would they start to collaborate?

Contact: Sebastian Risi (sebr@itu.dk)

32 of 45

REPURPOSE Project: Repurposing Existing Video Games for Citizen Science

  • Instead of designing games for a specific scientific challenge, can we develop a general interface that allows us to�repurpose already existing video games for crowdsourcing
  • Imagine the millions of players of popular video games contributing towards solving important scientific problems or complex computations while playing their favorite game.

Contact: Sebastian Risi (sebr@itu.dk)

33 of 45

Optimal Number of Bio-Inspired Learning Rules for Adaptive Neural Networks

Extend algorithm for evolving plastic neural networks to find the optimal number of unique plasticity rules.

Plasticity rules make neural networks more adaptive.

Plasticity rules can be optimized to control multiple network connections.

Can we through evolution find

the perfect number of unique

plasticity rules that enables

optimal adaptiveness of

bio-inspired neural networks?

Research Question:

Contact: Joachim Winther Pedersen jwin@itu.dk

34 of 45

Evolving computational graph topologies

Exploring the how agent adaptation is facilitated by cognition encoded in graph structures

Contact: Djordje- djgr@itu.dk

How does sparsely connected neural-network allows agent to continually adapt to novel stimuli?

  • Reinforcement learning
  • Graph neural networks
  • Evolutionary algorithms
  • Graph cellular automata

35 of 45

Gaining insights with Agent-based models (ABM)

Exploring the how multi-agent simulations can help us learn about collective outcomes

Contact: Djordje- djgr@itu.dk

Can reinforcement-learning-based improve ABMs?

  • Reinforcement learning
  • Multi-agent systems

Build products informed by emergent behavior in ABMs?

  • Reinforcement learning
  • Multi-agent systems

36 of 45

Make a GPU-accelerated Minecraft clone

Contact: Djordje- djgr@itu.dk

  • Minecraft is a game that is popular among the AI researchers
  • Standard Minecraft implementation runs sequentially on CPU.
  • Create an impact on the AI field by making a Minecraft clone that runs on GPUs

37 of 45

Language Cellular Automata

Combining Neural Cellular Automata models with Language models to get a language interface to cellular automata models

#NCA #LLMs #Generative

Contact: Elias Najarro, Eleni Nisioyi,

Milton Montero

enaj@itu.dk, enis@itu.dk, mlle@itu.dk

38 of 45

Climbing generator

Build a generative model for climbing routes & boulder problems using a combination of 3D pose estimation and reinforcement learning simulations

#RL #Generative models

Contact: Elias Najarro, Milton Montero

enaj@itu.dk, mlle@itu.dk

39 of 45

Open-endedness in Minecraft

Running evolutionary simulations with LLM agents in Minecraft survival mode to study the emergence of open-ended strategies

#Open-Endedness #LLMs

Contact: Elias Najarro

enaj@itu.dk

40 of 45

Emergence in Artificial Life

Investigating the emergence of artificial life forms known as virtual creatures using self-organisation and GPU accelerated simulations

#ALife #Emergence

Contact: Elias Najarro

enaj@itu.dk

41 of 45

Designing patterns to fool face recognition systems

The goal is combine use AI/ML methods to design patterns that can fool open source face detection and face recognition models and test them under real world conditions by painting/embroidering them.

#Design #Critical Tech #Optimisation

Contact: Elias Najarro

enaj@itu.dk

42 of 45

Drawing with Nature

Explore the expressivity and constraints of natural systems as generative systems

#Design #Generative

Contact: Elias Najarro

enaj@itu.dk

References: “Modern Evolution Strategies for Creativity: Fitting Concrete Images and Abstract Concepts”. Tian and Ha, 2022.

The Nature of Code: Simulating Natural Systems, Daniel Shiffman

43 of 45

Modelling neural growth

Build a generative model of how biological brains grow based on datasets from different species (such as C. elegans, fruit-fly, humans)

We will use recent generative models of graphs, such as Digress and Neural Developmental Programs, aiming at capturing

emergent features of neural growth

(related to structure and potentially function)

that simpler models cannot account for

(for example they often ignore neurogenesis

and do not scale)

Contact: Eleni Nisioti enis@itu.dk

Erwan Plantec erpl@itu.dk

44 of 45

Self-organising systems for robust visual perception

Explore the use self-organising systems such as Neural Cellular Automata for visual perception tasks such as image segmentation and representation of objects and scenes.

Contact: Milton Montero mlle@itu.dk

45 of 45

Self-organising systems for solving the ARC challenge

Explore the use self-organising systems such as Neural Cellular Automata for pattern generation following predefined rules as presented in the Abstract Reasoning Corpus challenge.

Contact: Milton Montero mlle@itu.dk