Object detectors like Mask R-CNN and YOLO are very good at locating and classifying known objects within images. In order to be successful though, these detectors generally need to be explicitly trained on a fixed number of classes with many (tens of) thousands of input samples. While the detectors are easy to use and train, obtaining and/or curating the appropriate training dataset for a particular set of objects is generally prohibitive and time-consuming. Open vocabulary object detectors can overcome this need for specialised training datasets, by locating objects in the image and classifying them with the use of language models. This allows novel objects to be detected simply by specification of their class in the form of (possibly descriptive) text. This thesis topic aims to construct an open vocabulary object detector, possibly based on YOLOv8 [1] and CLIP [2], while trying to learn from design decisions made by ViLD [3], CORA [4], and POMP [5].

Useful skills:

• Expert Python 3 programming skills (including experience working on/with large open-source libraries)
• Extensive experience using PyTorch and related libraries
• Interest in self-driven/independent study where you can implement and academically publish your ideas

Reference:

[1] YOLOv8: https://github.com/ultralytics/ultralytics
[2] CLIP: https://github.com/openai/CLIP
[3] ViLD: https://github.com/tensorflow/tpu/tree/master/models/official/detection/projects/vild
[4] CORA: https://github.com/tgxs002/CORA
[5] POMP: https://github.com/amazon-science/prompt-pretraining

Contact:

Dr. Philipp Allgeuer, Prof. Dr. Stefan Wermter

Recent works have made astounding progress toward generative models that can produce high quality 3D textured models. One example of such a model is Get3D [1], which is able to synthesise new models of a particular kind (whatever kind it was trained on, e.g. cars) based on a latent space input. Get3D was trained on subsets of the ShapeNet dataset [2], which provides many CAD models of different object classes. The task of this thesis is to extend the generative pipeline, ideally starting from the pretrained Get3D model, to allow image prompting, potentially first based on ShapeNet and then later based on images, e.g. from the ApolloCar3D dataset [3]. The aim is that a cropped and background-removed photo of a car can ultimately be used to generate a closely matching 3D textured car model, as well as estimate the orientation/viewpoint of the car in the photo.

Useful skills:

• Expert Python 3 programming skills (including experience working on/with large open-source libraries)
• Extensive experience using PyTorch and related libraries
• Interest in self-driven/independent study where you can implement and academically publish your ideas

Reference:

[1] Get3D / Image: https://github.com/nv-tlabs/GET3D
[2] ShapeNet: https://shapenet.org
[3] ApolloCar3D: https://apolloscape.auto/car_instance.html

Contact:

Dr. Philipp Allgeuer, Prof. Dr. Stefan Wermter

Large language models are increasingly used to write robot control programs rather than to emit action sequences directly. Early code generation approaches such as Code as Policies synthesize a single program from an instruction and execute it open loop, which is brittle once perception is noisy or the scene deviates from what the prompt described. Agentic coding changes the setting: the program becomes an artifact that an agent authors, executes, inspects, and repairs across many iterations, with execution feedback rather than human review closing the loop. Graph-as-Policy (GaP) is a recent example of this idea in robotics. An orchestration agent partitions a task into skill-aware segments, skill agents synthesize subgraphs from a modular skill library, and the wired, type-checked computation graph is refined through repeated rehearsal in simulation before running unchanged on the physical robot. In our own work, LLM+MAP shows that grounding an LLM in PDDL yields verifiable, temporally consistent long-horizon plans for the NICOL humanoid, but the step from a symbolic action to executable robot code is still written by hand. This topic investigates agentic coding as the missing interface between symbolic planning and execution. The student will connect a coding agent to the NICOL skill and ROS interfaces and let it write, run, and revise manipulation programs autonomously, using simulation rollouts and runtime feedback as the learning signal.

Possible directions include:

• Program representation. Compare flat Python scripts, typed computation graphs, and PDDL-grounded skill compositions with respect to interpretability, verifiability, and repairability.
• Self-improvement loop. Localize failures to individual skill calls from contact, state, and vision feedback, and verify candidate edits before they are deployed.
• Persistent skill library. Let the agent accumulate and refactor reusable skills across tasks, so that later tasks start from previously validated code.
• Sim-to-real transfer. Deploy refined programs on the physical NICOL robot for long-horizon bimanual tasks, including asynchronous ones where both arms act under timing constraints.
• Evaluation covers success rate under object pose and geometry variation, the number of agent iterations required to reach a stable program, and robustness compared with one-shot code generation and with LLM+MAP as a symbolic planning baseline.

Useful skills:

• Python programming, including practical experience with LLM APIs, tool calling, and agent frameworks such as Claude Code, Codex, or MCP-based harnesses
• Familiarity with ROS and simulation environments is helpful but can be learned during the project
• Interest in self-driven, independent study where you can implement and academically publish your ideas

Literature:

[1] Chen, Kaiyuan, et al. "GaP: A Graph-as-Policy Multi-Agent Self-Learning Harness for Variational Automation Tasks." arXiv:2607.05369 (2026)

[2] Chu, Kun, Xufeng Zhao, Cornelius Weber, and Stefan Wermter. "LLM+MAP: Bimanual Robot Task Planning using Large Language Models and Planning Domain Definition Language." arXiv:2503.17309 (2025)

[3] Liang, Jacky, et al. "Code as Policies: Language Model Programs for Embodied Control." ICRA 2023

[4] Ma, Yecheng Jason, et al. "Eureka: Human-Level Reward Design via Coding Large Language Models." ICLR 2024

Contact:

Kun Chu, Prof. Dr. Stefan Wermter

Transformers are currently state-of-the-art models for natural language processing and they have started to make a good impression in the computer vision field as well [1]. This research project aims at designing a novel architecture for exploiting the capabilities of attention-based models such as transformer networks and provide a robust methodology for improving human-robot interaction through the recognition of gestures as a way of communication. Here at WTM we have developed Snapture [2], an architecture that enables learning static and dynamic gestures. This architecture is based on convolutional neural networks and LSTMs. Stepping on this research we want to move now towards exploring the potential of transformer architectures. One possible way to exploit a transformer architecture for gesture recognition is by pairing it with a strong feature extractor such as MediaPipe [3], a general purpose machine learning open source project developed by Google with robust pose estimation capabilities

Goals:

• Design a novel deep neural network architecture based on transformers for gesture recognition
• Implement, train and evaluate the new architecture using a standard benchmark for gesture recognition
• Possibly create and make publicly available a new dataset for gesture recognition in the context of human-robot communication

Useful skills:

• Proficiency in Python coding
• Interest in human-robot communication
• Experience programming and training deep neural networks

Reference:

[1] Z. Cao , Y. Li, B-S. Shin. "Content-Adaptive and Attention-Based Network for Hand Gesture Recognition" [PDF]
[2] H. Ali, D. Jirak, S. Wermter (2022) "Snapture - A Novel Neural Architecture for Combined Static and Dynamic Hand Gesture Recognition" [PDF]
[3] MediaPipe

Contact:

Hassan Ali, Prof. Dr. Stefan Wermter

The ability to navigate through a semi-dynamic environment is an important skill for a humanoid robot. This skill extends its range and field of applicability for human-robot interaction scenarios. In recent work, we have fitted a 2D lidar sensor [1] to one of our Pepper robots [2], and established localisation and navigation abilities within simple office scenarios. The assembled navigation stack uses Adaptive Monte Carlo Localisation (AMCL) for localisation [3], and a mix of global and local path planners for navigation. The use of a neural method for the purposes of navigation is to be investigated, possibly inspired for instance by Neural SLAM [4], or neural lidar odometry methods such as CAE-LO [5], PSF-LO [6], or the probabilistic trajectory estimator used in [7].

As a reference navigation approach, a new and more robust localisation and navigation stack should be developed, which applies a Simultaneous Localisation and Mapping (SLAM) method, and concurrently registers the output thereof against a known 2D reference map. The navigation strategy needs to allow path planning to targets that are outside the current SLAM map, but within the more comprehensive reference map. As an extension goal, any need for manual localisation initialisation should be alleviated using a global matching strategy between local laser scans and the reference map. The performance of the final navigation stack should be experimentally compared to the existing one.

Goals:

• Apply and tune a SLAM method for 2D robot localisation on the Pepper robot
• Implement a path planning/navigation strategy that efficiently utilises the omnidirectional capabilities of the Pepper robot in order to drive to a global pose goal while avoiding obstacles
• Possibly develop a robust global localisation initialisation strategy

Useful skills:

• High proficiency in object-oriented C++ and Python coding
• Prior experience with the Robot Operating System (ROS) middleware and PyTorch
• Interest in navigation and mapping using mobile robots
• Experience working with Raspberry Pi and/or robotic systems

Reference:

[1] YDLidar G2 Sensor Datasheet
[2] SoftBank Robotics Pepper robot
[3] D. Fox, W. Burgard, F. Dellaert, and S. Thrun (1999), “Monte carlo localization: Efficient position estimation for mobile robots” [PDF]
[4] J. Zhang, L. Tai, M. Liu, J. Boedecker, and W. Burgard (2020), "Neural SLAM: Learning to Explore with External Memory" [PDF]
[5] D. Yin, Q. Zhang, J. Liu, X. Liang, Y. Wang, J. Maanpää, H. Ma, J. Hyyppä, and R. Chen (2020), "CAE-LO: LiDAR Odometry Leveraging Fully Unsupervised Convolutional Auto-Encoder Based Interest Point Detection and Feature Description" [PDF]
[6] G. Chen, B. Wang, X. Wang, H. Deng, B. Wang, and S. Zhang (2021), "PSF-LO: Parameterized Semantic Features Based Lidar Odometry" [PDF]
[7] D. Yoon, H. Zhang, M. Gridseth, H. Thomas, T. Barfoot (2021), "Unsupervised Learning of Lidar Features for Use in a Probabilistic Trajectory Estimator" [PDF]

Contact:

Dr. Philipp Allgeuer, Prof. Dr. Stefan Wermter

World models are currently studied along two largely separate lines. In model-based reinforcement learning, an agent learns latent dynamics from its own interaction and then improves a policy by imagining rollouts or by planning over short horizons, which yields strong sample efficiency and genuine closed-loop control, but the resulting models are usually trained per environment and generalize poorly beyond the tasks and objects they were trained on. In vision-language-action modeling, a policy inherits broad semantic and visual generalization from large-scale pretraining, yet remains reactive: it maps observation and instruction to the next action chunk without predicting the consequences, so it degrades under object pose and geometry variation and cannot distinguish a promising action from a doomed one before executing it. A predictive model of the environment is the component that each line has and the other lacks. This topic investigates world models as the shared substrate between the two. The platform can be the NICOL humanoid for bimanual tabletop manipulation, or a simulated humanoid for whole-body control, depending on the student's interest. Three questions are central:

• 1. What should the model predict? Candidates range from pixel-level video prediction, to latent dynamics in a compact recurrent or transformer state space, to physics-structured representations that encode energy, momentum, or contact explicitly. Each choice trades rollout fidelity against horizon length, inference cost, and how well the model transfers across tasks and embodiments. The student will implement at least two and compare them on held-out interaction data.
• 2. How is the model used to obtain a policy? On the reinforcement learning side, the model can train a policy in imagination or serve as the dynamics for latent model predictive control. On the VLA side, the same model can rank and refine action chunks proposed by a pretrained policy, supply an auxiliary predictive objective during finetuning, or flag the divergence between predicted and observed outcomes as a trigger for recovery. Whether one model can serve both roles, and what it costs at runtime, is an open question the project addresses directly.
• 3. Where does the data come from? Interaction data is expensive on a physical robot, while action-free video is abundant. The student will examine how far a world model can be pretrained from passive observation, for example through latent action learning, and how much on-robot or in-simulation interaction is then needed to make it accurate enough for control.

Evaluation covers prediction accuracy over increasing horizons, sample efficiency on the reinforcement learning side, task success and robustness to disturbances on the policy side, and the latency added per control step relative to a model-free baseline.

Useful skills:

• Python and practical experience with deep learning frameworks, ideally including sequence models, diffusion policies, or reinforcement learning libraries
• Interest in reinforcement learning, model predictive control, or physics-informed machine learning
• Willingness to work with robot simulation and, if the project goes well, with the physical robot
• Interest in self-driven, independent study where you can implement and academically publish your ideas

References:

[1] Hou et al. "World Model for Robot Learning: A Comprehensive Survey." arXiv:2605.00080 (2026)

[2] Hafner, Danijar, et al. "Mastering Diverse Domains through World Models." Nature, 2025

[3] Hansen, Nicklas, Hao Su, and Xiaolong Wang. "TD-MPC2: Scalable, Robust World Models for Continuous Control." ICLR 2024

[4] Ye et al. "Latent Action Pretraining from Videos (LAPA)." ICLR 2025

[5] Tian et al. "Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (Seer)." ICLR 2025

[6] Tang et al. "DreamSAC: Learning Hamiltonian World Models via Symmetry Exploration." CVPR 2026

Contact:

Kun Chu, Prof. Dr. Stefan Wermter

Human pose estimation is a field that has seen many advances in the past years, and is important in human-robot interaction scenarios for a greater robot situational awareness. We wish to focus on the detection and wholistic understanding, from monocular images, of human body parts and keypoints in both 2D and 3D. This is a broad topic that offers a multitude of research directions for a prospective thesis, which in distant future work would seek to be united. The following individual possible goals are considered:

1. Can a feature backbone successfully be shared for human bounding box detection and 2D keypoint detection?
   Bounding box detection and keypoint detection are both time intensive processes, and performing both concurrently in a single network offers computational savings, as well as the sharing of knowledge, i.e. the visual understanding of human body parts [1] [2] [3]. Refer to new research in this direction [9].

2. Can 3D keypoint detection performance be improved by including the backbone feature maps as further inputs?
   3D keypoint estimators typically only receive 2D keypoint detection coordinates as input [4], thereby not allowing the network to learn from any visual context that can provide hints regarding limb depth and/or positional ambiguity.

3. Can 2D keypoint detection performance in the wild be improved by training on multiple united datasets?
   Single datasets can lack the complete diversity required to train networks that are also effective in the wild [5].

4. Can a more wholistic understanding of the appearance of the human body be learnt by concurrently training multiple related tasks?
   Examples of candidate tasks in addition to 2D keypoint estimation include instance segmentation, body part segmentation [6], dense pose estimation [7], and more if considering 3D pose estimation [8]. Related idea [10].

5. Can temporal information be used to improve 2D keypoint detection robustness for video sequences?
   It is common for 3D keypoint estimators to use sequences of multiple frames in order to improve their estimation quality in the face of ambiguities [4]. Can 2D keypoint detectors also benefit in order to help deal with body part occlusions?

Useful skills:

• Expert Python 3 programming skills (including experience working on/with large open-source libraries)
• Extensive experience using PyTorch and related libraries
• Interest in self-driven/independent study where you can implement and academically publish your ideas

Reference:

[1] Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, "YOLOX: Exceeding YOLO Series in 2021", arXiv preprint arXiv:2107.08430, 2021. [PDF]
[2] D. Bolya, C. Zhou, F. Xiao, Y. Lee, "YOLACT++: Better Real-time Instance Segmentation", arXiv preprint arXiv:1912.06218, 2020. [PDF]
[3] Z. Huang, L. Huang, Y. Gong, C. Huang, X. Wang, "Mask Scoring R-CNN", arXiv preprint arXiv:1903.00241, 2019. [PDF]
[4] D. Pavllo, C. Feichtenhofer, D. Grangier, M. Auli, "3D human pose estimation in video with temporal convolutions and semi-supervised training", arXiv preprint arXiv:1811.11742, 2019. [PDF]
[5] J. Lambert, Z. Liu, O. Szener, J. Hays, V. Koltun, "MSeg: A Composite Dataset for Multi-domain Semantic Segmentation", arXiv preprint arXiv:2112.13762, 2021. [PDF]
[6] W. Wang, Z. Zhang, S. Qi, J. Shen, Y. Pang, L. Shao, "Learning Compositional Neural Information Fusion for Human Parsing", arXiv preprint arXiv:2001.06804, 2020. [PDF]
[7] R. Güler, N. Neverova, I. Kokkinos, "DensePose: Dense Human Pose Estimation In The Wild", arXiv preprint arXiv:1802.00434, 2018. [PDF]
[8] A. Mertan, D. Duff, G. Unal, "Single Image Depth Estimation: An Overview", arXiv preprint arXiv:2104.06456, 2021. [PDF]
[9] D. Maji, S. Nagori, M. Mathew, D. Poddar, "YOLO-Pose: Enhancing YOLO for Multi Person Pose Estimation Using Object Keypoint Similarity Loss", arXiv preprint arXiv:2204.06806, 2022. [PDF]
[10] C. Wang, I. Yeh, H. Liao, "You Only Learn One Representation: Unified Network for Multiple Tasks", arXiv preprint arXiv:2105.04206, 2021. [PDF]

Contact:

Dr. Philipp Allgeuer, Prof. Dr. Stefan Wermter

In human-human communication, gestures are essential components of non-verbal communication and have been identified as social cues in human-robot interaction (HRI). Co-speech gestures are delivered with speech and can enhance specific word meanings. The examples for co-speech gestures range from static poses like "OK" to dynamic hand movements while giving a public talk. While the recognition of those gesture forms is inevitably necessary to foster natural HRI scenarios, most research in gesture recognition has focused only on a separate aspect of gestures, either hand pose or dynamic gesture recognition. We recently tackled the integration of both modalities introducing a novel neural network architecture called "Snapture" [1]. Snapture is a modular framework that learns both a specific finger pose and the arm movements involved in the gesture expression. Snapture is based on convolutional neural networks (CNN), which learn hand poses and long short-term memory networks (LSTM) to model the temporal sequences. We showed the robustness of the architecture using the Montalbano co-speech benchmark dataset [2]. The Snapture framework offers a multitude of extensions for a prospective Master thesis: 

• Improvement of the threshold-based activation of the static hand recognition channel for the flexible detection of hand poses in a sequence.

• Analysis of the neuronal network activations (e.g. grad-Cam or embeddings) and integration of attention mechanisms to understand the working principles of the Snapture architecture.

• Integration of facial expressions in alignment with co-speech gestures to add an affective dimension to the gesture expression context.

• Optional (depending on the Covid19 situation): Implementation of HRI co-speech scenarios with a humanoid robot.

The Snapture framework will be made available at the start of the Master thesis. We require a profound knowledge of artificial neural networks, their learning principles, and network performance evaluation. Programming experience in Python is mandatory. Experiences with gesture recognition both in human-human and human-robot interaction is a plus.

Useful skills:

• Profound knowledge of artificial neural networks

• Programming experience with Python

• Interest in gesture recognition

Reference:

[1] Hassan Ali: Snapture-A Hybrid Hand Gesture Recognition System, 2021

[2] https://chalearnlap.cvc.uab.es/dataset/13/description/

Contact:

https://www.inf.uni-hamburg.de/en/inst/ab/wtm/people/ali.html

Hassan Ali, Dr. rer. nat. Doreen Jirak, Prof. Dr. Stefan Wermter

In recent work [1], we have tested the capability of a robot to tell about what it is doing. It learnt to produce phrases like "push red pylon", even if it has never before seen a red pylon, but only pylons of other colors and other red objects. Moreover, additional irrelevant object within the robot's field of view do not disturb much. However, the performance drops drastically, if there are not many color and object combinations in the dataset (while keeping the dataset of constant size). In this thesis, we want to tackle this difficult case for generalization. We will investigate whether unsupervised pretraining (either with a small number of color and object combinations, or with a larger number, but always without labels) would help. We plan also to consider importing models for preprocessing that are pretrained on large vision-language datasets, such as ViLBERT [2]. We will validate the success of the extended model on the difficult multi-object cases in our earlier experiments [1].

Goals:

• Develop a model which describes an action performed in a scene on one of several objects

Useful skills:

• Neural networks
• Interest in 3D robot simulator CoppeliaSim

References:

[1] Generalization in Multimodal Language Learning from Simulation
Aaron Eisermann, Jae Hee Lee, Cornelius Weber, Stefan Wermter, IJCNN 2021
[2] ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks
Lu et al., NeurIPS 2019

Contact:

Jae Hee Lee, Cornelius Weber, Prof. Dr. Stefan Wermter

The power of computational natural language processing can only be unleashed when meaningful dialogues can be generated. The aim of this project is to implement a minimal model of dialogue about a visual scene. On the example of a "Guess Which" game we will implement a dialogue about several objects in a visual scene: an "OracleBot" chooses one designated object, and a "LearnerBot" needs to guess which object is chosen, which it can only infer via a questioning strategy, such as asking about which color, size, shape, etc. the object has that the OracleBot has in mind. The LearnerBot model uses a mixed strategy of reinforcement learning (to learn the question strategy) and supervised learning (for the answer). The model is based on a model of sequential visual processing that delivers an answer from accumulated visual impressions [1], but we will transfer it to the visual-linguistic domain [2].

Goals:

• The development of a neural network model for the LearnerBot in the GuessWhich game.

Useful skills:

• Programming experience with Python
• Familiarity with deep learning and recurrent neural networks
• Interest in natural language processing
• Experience with image processing or visual simulation environments

References:

[1] V. Mnih, N. Heess, A. Graves, K. Kavukcuoglu (NeurIPS 2014), Recurrent Models of Visual Attention

[2] End-to-end Optimization of Goal-driven and Visually Grounded Dialogue Systems

Contact:

Dennis Becker, Cornelius Weber, Prof. Dr. Stefan Wermter

In the context of assistive robots, verbal user instructions to a robot often refer to objects that are visible in the immediate surround. Robot speech recognition could therefore benefit from visual context by biasing recognition towards words that denote seen objects. Task is to create such an audio-visual model and a methodology to train it.

Goals:

• Develop a novel speech recognition architecture with additional visual input
• Find a dataset and develop a training methodology
• Evaluate the trained model

Useful skills:

• Programming experience with Python
• Familiarity with deep learning
• Interest in speech recognition and vision processing

References:

[1] https://pjreddie.com/darknet/yolo/

[2] https://missinglink.ai/guides/tensorflow/tensorflow-speech-recognition-two-quick-tutorials/

Contact:

Cornelius Weber, Marie Sophie Bauer, Prof. Dr. Stefan Wermter

“Explainable Agent” refers to autonomous agents that can explain their actions and the reasons leading to their decisions. This will incite their users to understand their capabilities and limits, thereby improving the levels of trust and safety, and avoiding failures. The Neural State Machine (NSM) is an explainable agent that bridges the gap between neural and symbolic AI by integrating their complementary strengths for the task of visual reasoning [1]. Given an image, the NSM first predicts a probabilistic semantic graph that represents the underlying semantics and serves as a structured world model. The structured world model is then used to perform sequential reasoning over the graph. The NSM provides spatial relational reasoning through a probabilistic scene graph. However, the Graph R-CNN used in the NSM is a black-box model. Therefore, the users may not be able to understand the reasons and causes for the NSM outputs. In addition, the NSM is not able to link meaningful transformations of objects or entities of the semantic graph over time. The objective of this project is, therefore, to develop an explainable NSM with Spatio-temporal reasoning.

spatial relational reasoning [1]

temporal relational reasoning (http://relation.csail.mit.edu/)

Goal

• To improve the effectiveness of explanations by improving model interpretability of the NSM [2].
• To enhance the NSM with Temporal Relational Reasoning capability.

Useful skills:

• Programming experience, preferably Python and frameworks such as PyTorch, Tensorflow and Keras,

• Knowledge in deep neural networks.

Reference

[1] Hudson, D. A. & Manning, C. D. (2019). Learning by Abstraction: The Neural State Machine. CoRR, abs/1907.03950.

[2] Stefan Wermter (2000). Knowledge Extraction from Transducer Neural Networks. Journal of Applied Intelligence, Vol. 12, pp. 27-42. https://www2.informatik.uni-hamburg.de/wtm/publications/2000/Wer00b/wermter.pdf

Contact:

Prof. Dr. Loo Chu Kiong, Dr. Cornelius Weber, Dr. Matthias Kerzel, Prof. Dr. Stefan Wermter

In humans and animals, responses to events in the environment depend critically on visual and auditory input integration. Our neuro-computational models of multimodal integration are embedded in robotic hardware and trained in controllable virtual reality robotic environments. The goal of this thesis is to develop a virtual environment that realizes a multi-agent conversation with realistic turn taking behaviour. The agents should be able to utilize both auditory cues (speaking up, interrupting each other) as well as visual cues (motion, gestures, head orientation) to signal turn taking. This task encompasses the design of individual virtual agents capable of expressing verbal and visual turn taking signals in the virtual robot experimentation platform V-REP and the development of a planning system that schedules turn taking for all agents to create a collaborative or conference like scenario. This thesis offer is part of the research project: “Vision- and action-embodied language learning”.

Goals:

• Realization of a virtual reality multi-agent conversation scenario
• Development of a framework for realistic conversational turn taking behaviour of virtual agents using visual and auditory cues.

Useful skills:

• Interest in linguistics, especially with the topic of turn taking.
• Experience with the virtual robot experimentation platform V-REP.
• Good programming skills in Python.

Contact:

Dr. Matthias Kerzel, Prof. Dr. Stefan Wermter

Our humanoid robot NICO, once derived from a soccer robot, lost its walking ability, since the classical control algorithm cannot handle the added weight of its hardware extensions on head and arms. Later, a bio-inspired control algorithm based on a central pattern generator produced, however slow, walking of NICO in the CoppeliaSim simulator [1].

Goals:

• Generate natural walking of NICO. This shall be feasible due to (i) the new powerful TD-MPC reinforcement learning algorithm, which has been shown to control a humanoid robot [2], and (ii) the new fast MuJoCo physics engine in CoppeliaSim, which allows more efficient training.

Useful skills:

• Interest in reinforcement learning, first experiences are a plus
• Ability to adjust settings in CoppeliaSim to counter unwanted effects, such as slippage
• Understanding of physical behaviour, so to adjust the cost/reward function to lead to natural walking

References:

[1] Hierarchical Control for Bipedal Locomotion using Central Pattern Generators and Neural Networks
https://www2.informatik.uni-hamburg.de/wtm/publications/2019/AMW19/auddy_icdl_2019.pdf
[2] Temporal Difference Learning for Model Predictive Control
https://nicklashansen.github.io/td-mpc/

Contact:

Mostafa Kotb, Cornelius Weber

Children learn to the know the environment but also to think in and express natural language by direct interaction with objects and other strange things in the environment. While doing this they make use of effective strategies from clumsily touching objects up to exploring it's affordances by manipulating it.

In this project we want to development these strategies for our NICO robot as a 3D-model in the virtual environment V-REP as well as our real robot in the lab. Furthermore we want to measure the perceived effect on the robot and on the objects and let the robot learn about object features over a longer time of interaction. As a long term goal we plan to employ the result of the project in our research on embodied language acquisition through automated linguistic instructions during robot object interaction. For this several students could collaborate e.g. working on visually tracking the manipulated objects and forming representations for perceived interactions in architectures in GWR and MTRNN architectures.

Goals:

• Study into motor babling in infants and develop motion strategies for the NICO,
• Develop an interface for the motion characteristics,
• For a Master Thesis: Learn effective interaction with different objects in a neural architecture, e.g. recurrent and self-organising.

Useful skills:

• Interest in robotics, motion, haptics, and learning, Affinity for child development and cognitive psychology,
• Programming skills in Python,
• Experience in V-Rep or other simulators, and with the NICO robot.

Contact:

Dr. Matthias Kerzel, Dr. Cornelius Weber, Prof. Dr. Stefan Wermter

Our NICO humanoid robot, when seated at a table, can only reach a small area in front of it, such as to grasp an object, because of its short arms. In order to extend its reaching range, in this project we plan to involve movements of its thighs, letting NICO tilt its upper body forward, thereby expanding its reaching horizon. Since the head with camera will move together with the torso, visuo-motor coordination will become more challenging compared to the normal case where the camera is kept in a fixed position.

Useful skills:

• Interest in reinforcement learning (RL), first experiences are a plus. We will aim for a sample-efficient RL algorithm (e.g. [1]), possibly aided by self- or unsupervised representation learning.
• Interest in 3D robot simulators (e.g. [2]), to model the NICO scenario.

References:

[1] Temporal Difference Learning for Model Predictive Control
https://nicklashansen.github.io/td-mpc/
[2] CoppeliaSim
https://www.coppeliarobotics.com/

Contact:

Mostafa Kotb, Cornelius Weber

In this exciting project we look at neuroscience-inspired architectures for controlling the behavior of a robot. While traditional robots have often been preprogrammed, our new approach will focus on learning  robots which will be based on some neuroscience evidence. These navigation and movement concepts are transferred and further developed on a Nao robot which has some speech and vision capabilities.  The NeuroBot will  learn to associate actions with words and pointing gestures. We want to restrict the fixed manual programming of the robot and emphasize the adaptive autonomous learning in neural networks in combination with  restricted instructions via words and simple pointing. Some of the previous and current robot platforms available are shown at http://www2.informatik.uni-hamburg.de/wtm/neurobots. Students who participated in our project "Human Robot Interaction" can also suggest a further topic for their theses.

Goals:

• Learning navigation from simple multimodal input
• Robotic vision: Object recognition and object manipulation
• Implementation of neural network algorithms for speech and pointing instructions for NeuroBot
• At later stage: vision and speech capabilities for the NAO

Requirements:

• Programming skills: C/C++, Python
• At least basic knowledge in neural network algorithms and natural language processing
• Willingness to work in robotic environment

The thesis can be written either in German or English. All topics could be tailored to be at a bachelor, master level, or phd student level. If you are interested contact us for discussion:

Contact:

Prof. Dr. Stefan Wermter, Dr. Cornelius Weber