Embedded Files
We offer interesting topics for student theses and projects as well as paid internships. Feel free to contact me for more information.
Other topics than listed below can be defined upon request – simply drop by at KN-E211 or write an email to matej.hoffmann [guess-what] fel.cvut.cz
List of currently open topics supervised by Matej Hoffmann with a formal description is available here: https://hub.fel.cvut.cz/topics/semestral_projects (filter projects supervised by Matěj Hoffmann).
Below includes is a list of teasers for some of the topics, including smaller projects for internships etc.
Resilient robots with sensitive skin on the whole body
The adaptivity and resilience of biological systems are still unparalleled by modern machines. Animals utilize numerous parallel sensory-motor loops and exploit whole-body contacts with their environment to navigate and manipulate the world safely. This project aims to develop biologically inspired controllers for robotic platforms that can perform exploration and manipulation tasks with extreme robustness.
The student will utilize robots equipped with rich whole-body sensing modalities—such as electronic skin, joint torque sensors, or motor current feedback (available on platforms like the UR10e, KUKA LBR iiwa, or the humanoid iCub)—to build systems that learn from and react to explorative contacts with the environment. The ultimate goal is to demonstrate that the robot can successfully explore environments and manipulate objects even under dramatic disruptions—such as manipulated encoder values, blocked joints, or distorted visual inputs—by relying on alternative, bio-inspired sensory-motor strategies that fuse diverse haptic and proprioceptive signals.
AI tools for developmental science
Detecting arousal and active exploration from videos of children - Link to hub.fel.cvut.cz
Detecting posture and movement dynamics of infants around motor developmental milestones with automated pose estimation - Link to hub.fel.cvut.cz
Young infants spend most of their wake time on their backs, producing seemingly random movements. Using a large dataset of video recordings of infants, our goal is to identify early signs of active exploration and emerging goal-directedness in spontaneous movement. Next to posture and movements automatically extracted from videos, the goal of this project is to explore the possibility of using other markers that could be extracted from images: changes in pupil size, blinking, and emotional states from faces or heart rate from the skin surface.
Object handover between human and robot
Object handover is a common task in human life. It can be described as an interaction between two agents where one agent hands an object to the second agent. In the context of robotics, the agents can be human and robot, or two robots. This project aims to create a pipeline that allows smooth and safe object handover between a human and a robot (humanoid robot iCub or robotic arm) or between two robots. The intermediate steps consist of detecting and tracking objects in the scene, controlling the robot, and reacting to obstacles or other unpredictable situations. Thus, it requires a combination of robotics, computer vision, and deep learning (possibly with LLMs).
Developmental robotics in a simulated infant
We offer several student projects using the Multimodal Infant Model (MIMo) simulator, an open-source MuJoCo-based simulator of an infant body with vision, proprioception, vestibular sensing, full-body touch, and muscle-like actuation. These projects use MIMo as a testbed for studying early sensorimotor development: how an embodied agent learns to reach, avoid, protect its body, respond to touch, generate reflex-like behavior, and acquire structured experience through interaction with objects and caregivers.
The projects range from implementing reinforcement-learning models of near-body action value (project on hub.fel.cvut.cz), through adding a biologically inspired infant reflex layer (project on hub.fel), to creating a simulated caregiver/mother who can physically interact with the baby robot (project on hub.fel), and developing motion retargeting tools that replay real infant movements in MIMo (project on hub fel). Together, these projects aim to connect real infant behavior, computational modeling, and developmental robotics. Each project can be adapted to a semester project, bachelor thesis, or master thesis depending on the student’s background and ambition. More detailed project descriptions are available below.
Instant Policy Learning for Robot Arms
Link to project on hub.fel.cvut.cz
The latest wave of robot learning models are trying to make robots which can carry out many different, diverse tasks. State-of-the-art methods like VLAs or world models require hundreds of thousands of episodes of data to train, which is costly to collect and store; the training process takes a long time and lots of compute power; and because these models are built on one big black-box network, analysis of the robot’s performance is difficult to explain. To tackle these problems, we are exploring alternative methods of learning task skills. In this project, we will deploy the Instant Policy paradigm on our robotic arms at CVUT FEL. This model is designed to learn the relative movements between only the robot’s gripper and the target object, and so the learning can be done in very few (<10) demonstrations.
The first part of the project will be to deploy Instant Policy, for a simple pick-and-place task, on our Kinova Gen-3 robot with the Robotiq 2F-85 gripper. Then, we will evaluate how the method functions, the accuracy of the model, the variability of the training process—how much it depends on the setup (video, lighting, the robot’s control parameters, etc.). The next part (optional) can be to try and extend the policy, so that the robot can learn the same skill with different objects, and then different skills.
Page updated
Google Sites
Report abuse