Alper Ahmetoglu

• Ahmetoglu, A., James, S., Allen, C., Lobel, S., Abel, D., Konidaris, G.D. (2025). Skill-Driven Neurosymbolic State Abstractions. Advances in Neural Information Processing Systems 38, 10750-10785. project page, poster
  Construct state abstractions compatible with a given set of abstract actions, to obtain a well-formed abstract MDP. We show that the Bellman equation suggests that abstract states should represent distributions over states in the ground MDP; we characterize the conditions under which the resulting process is Markov and approximately model-preserving, derive algorithms for constructing and planning with the abstract MDP.
• Kilic, B., Ahmetoglu, A., Ugur, E. (2025). Predictability-Based Curiosity-Guided Action Symbol Discovery. In 2025 IEEE International Conference on Development and Learning (ICDL) (pp. 1-6).
• Ugur, E., Ahmetoglu, A., Nagai, Y., Taniguchi, T., Saveriano, M., Oztop, E. (2025). Neuro-Symbolic Robotics. (preprint)
  A summary of neuro-symbolic robotics. We introduce four main categories based on the type of usage of neural and symbolic components.
• Ahmetoglu, A., Oztop, E., Ugur, E. (2025). Symbolic Manipulation Planning with Discovered Object and Relational Predicates. IEEE Robotics and Automation Letters (RA-L) 10(2), 1968-1975. project page, poster, video
  Translating a set of transitions encoded with deep nets as object and relational symbols into PDDL rules.
• Ahmetoglu, A., Celik, B., Oztop, E., Ugur, E. (2024). Discovering Predictive Relational Object Symbols with Symbolic Attentive Layers. IEEE Robotics and Automation Letters (RA-L) 9(2), 1977-1984. project page, video, code
  Differentiably discretized self-attention weights act as relational symbols between objects, giving a unified architecture, Relational DeepSym, in which the state, composed of objects, can be translated into a set of object and relational symbols that encode effects of actions. The composition of multiple objects forms a set of object and relational symbols which encodes how the environment changes with respect to actions.
• Celik, B., Ahmetoglu, A., Ugur, E., Oztop, E. (2023). Developmental Scaffolding with Large Language Models. In 2023 IEEE International Conference on Development and Learning (ICDL) (pp. 396-402).
  Guiding the exploration process via LLMs acting as a scaffolding agent.
• Ahmetoglu, A., Oztop, E., Ugur, E. (2022). Learning Multi-Object Symbols for Manipulation with Attentive Deep Effect Predictors. arXiv:2208.01021. video
  An increment over the previously proposed DeepSym architecture, in which a varying number of objects can be given as input and processed via self-attention layers. The composition of multiple objects forms a specific symbol which encodes how the environment changes with respect to actions.
  IEEE Best Paper Award at SIU 2023 (top ML conference in Turkey).
• Ahmetoglu, A., Seker, M.Y., Piater, J., Oztop, E., Ugur, E. (2022). DeepSym: Deep Symbol Generation and Rule Learning from Unsupervised Continuous Robot Interaction for Planning. Journal of Artificial Intelligence Research 75, 709-745. video, code
  An encoder-decoder network with symbolic bottleneck layer trained to predict the effects of actions given the state and action learns affordance-like symbols (either per object or for the global representation) that can be translated into PDDL for domain-independent planning.
• Ahmetoglu, A., Ugur, E., Asada, M., Oztop, E. (2022). High-level Features for Resource Economy and Fast Learning in Skill Transfer. Advanced Robotics 36(5-6), 291-303.
  Applying slow feature analysis (SFA) on top of a previously trained neural net provides a compact set of units that can be used for subsequent tasks.
• Seker, M.Y., Ahmetoglu, A., Nagai, Y., Asada, M., Oztop, E., Ugur, E. (2022). Imitation and Mirror Systems in Robots through Deep Modality Blending Networks. Neural Networks 146, 22-35.
  Reconstructing multiple modalities from each other, e.g., predicting the visual modality from proprioceptive input or vice versa, not only creates a correspondence between multiple modalities but also provides a representation that generalizes to out-of-distribution examples such as different grippers, colors, and objects, even though the system is only trained with a single instance of such properties.
• Gokay, D., Simsar, E., Atici, E., Ahmetoglu, A., Yuksel, A.E., Yanardag, P. (2021). Graph2Pix: A Graph-Based Image to Image Translation Framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops (pp. 2001-2010). project page, video, code
  An image translation framework from a graph of images.
• Aksoy, C., Ahmetoglu, A., Gungor, T. (2020). Hierarchical Multitask Learning Approach for BERT. arXiv:2011.04451.
  An analysis of BERT's downstream performance when trained with the auxiliary tasks at different layers.
• Ahmetoglu, A., Alpaydin, E. (2020). Hierarchical Mixture of Generators for Adversarial Learning. In 2020 25th International Conference on Pattern Recognition (ICPR) (pp. 316-323). project page, video, slides, code
  Training a hierarchical mixture of generators increases both the quality and coverage of generated samples in the context of GANs. The learned tree structure also allows for the interpretation of generators.
• Ahmetoglu, A., Irsoy, O., Alpaydin, E. (2018). Convolutional Soft Decision Trees. In Proceedings of the 27th International Conference on Artificial Neural Networks (ICANN) (pp. 134-141). slides
  
  A differentiable combination of a deep net with (cooperative) hierarchical mixtures of experts slightly increases the classification performance and provides interpretability of decisions.

Theses

Organized Workshops

• IROS 2026 Workshop on Bridging the Gap between Neural and Symbolic World Models for Robot Planning, Reasoning, and Action
• IROS 2022 Workshop on Lifelong Learning of High-level Cognitive and Reasoning Skills
• RSS 2021 Workshop on Declarative and Neurosymbolic Representations in Robot Learning and Control

Area Chair

• International Conference on Robotics and Automation (ICRA 2026)
• International Conference on Humanoid Robots (Humanoids 2024, 2025)
• International Conference on Development and Learning (ICDL 2025)