Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Tamás Bécsi has a total of 61 co-authors. Here are the top co-authors by number of shared publications:

  • Szilárd Aradi

    Department of Control for Transportation and Vehicle Systems, Faculty of Transportation Engineering and Vehicle Engineering, Budapest University of Technology and Economics, Budapest, Hungary

    58 shared publications
  • Péter Gáspár

    Hungarian Research Network (HUN-REN)

    18 shared publications
  • Bálint Kővári

    Department of Control for Transportation and Vehicle Systems, Faculty of Transportation Engineering and Vehicle Engineering, Budapest University of Technology and Economics, Budapest, Hungary

    16 shared publications
  • Olivér Törő

    Budapest University of Technology and Economics,Department of Control for Transportation and Vehicle Systems,Budapest,Hungary

    13 shared publications
  • Peter Gaspar

    Systems and Control Laboratory and Institute for Computer Science and Control, Budapest University of Technology and Economics, Budapest, Hungary

    11 shared publications
  • Adam Szabo

    Budapest Univ. of Technology and Economics,Department of Control for Transportation and Vehicle Systems,Budapest,Hungary

    6 shared publications

Latest Publications

  • Open Access

    Uncertainty aware stochastic sampling for efficient object detection

    by Csanád Levente Balogh, Bence Pap, Bálint Kővári, Tamás Bécsi

    The effectiveness of deep learning models is strongly influenced by the quality of training data. Traditional training approaches assume that all samples contribute equally to the learning process, leading to uniform data sampling. However, this assumption overlooks the substantial variation in informational content across samples. This paper introduces a novel and computationally lightweight data prioritization methodology for object detection that dynamically adjusts the sampling probability of training data according to its relevance during learning.The proposed Relative Detection Error (RDE) is a new temporal instability metric defined in a joint classification–regression setting, where classification corresponds to object category prediction and regression corresponds to bounding-box parameter estimation. By explicitly quantifying prediction variability over time, RDE identifies samples with higher learning value and guides an exploration-regularized stochastic sampling policy. As a result, the method improves both classification and localization accuracy while imposing minimal computational overhead and integrating seamlessly into standard training pipelines. The approach is validated using YOLO architectures on diverse datasets, demonstrating consistent improvements and strong generalization across models and domains. Experimental results show that prioritizing high-value samples yields higher F1 scores and mean Average Precision, alongside more efficient and stable convergence. The project source code is available at: https://kp-labs-bme.github.io/Object-Detection-Prioritization/

  • Open Access

    Physics-informed neural particle flow for the Bayesian update step

    by Domonkos Csuzdi, Tamás Bécsi, Olivér Törő


  • Open Access

    A Novel Sensor Model-Based Object Existence Probability Fusion for Automotive Track-to-Track Fusion

    by Jenő Bóka, László Lindenmaier, Szilárd Aradi, Tamás Bécsi

    Environment perception—and, in particular, multi-sensor fusion—plays a crucial role in ensuring the reliability of advanced driver assistance systems. Track-to-track fusion principles are widely adopted in automotive applications to combine object tracks generated by smart sensors. Within these systems, track management is a key component, as it strongly influences performance through track initiation and false-track suppression. State-of-the-art approaches address this problem by estimating and fusing the existence probabilities of tracks. Consequently, robust fusion of sensor-level existence estimates is essential for reliable track-to-track methods. In this paper, we extend our previous two-step hybrid existence probability fusion framework with probabilistic tracking and birth models. The proposed tracking model incorporates target velocity in addition to the sensor field of view, while the birth model captures the spatial context in which new objects are likely to emerge, enabling faster confirmation and improved false-track suppression. Experimental results obtained with a conventional radar–camera frontal perception setup demonstrate that integrating the proposed sensor models yields more than a 4% overall performance improvement, even within a standard one-step Dempster–Shafer belief fusion framework.

  • Open Access

    Physics-informed neural particle flow for the Bayesian update step

    by Domonkos Csuzdi, Bécsi Tamás, Olivér Törő

    The Bayesian update step poses significant computational challenges in high-dimensional nonlinear estimation. While log-homotopy particle flow filters offer an alternative to stochastic sampling, existing formulations usually yield stiff differential equations. Conversely, existing deep learning approximations typically treat the update as a black-box task or rely on asymptotic relaxation, neglecting the exact geometric structure of the finite-horizon probability transport. In this work, we propose a physics-informed neural particle flow, which is an amortized inference framework. To construct the flow, we couple the log-homotopy trajectory of the prior to posterior density function with the continuity equation describing the density evolution. This derivation yields a governing partial differential equation (PDE), referred to as the master PDE. By embedding this PDE as a physical constraint into the loss function, we train a neural network to approximate the transport velocity field. This approach enables purely unsupervised training, eliminating the need for ground-truth posterior samples. We demonstrate that the neural parameterization acts as an implicit regularizer, mitigating the numerical stiffness inherent to analytic flows and reducing online computational complexity. Experimental validation on multimodal benchmarks and a challenging nonlinear scenario confirms better mode coverage and robustness compared to state-of-the-art baselines.

  • Open Access

    Learning to Predict Constraints: Hybrid Neural-MPC Control Architectures for Real-Time Vehicle Path Tracking

    by Ákos M. Bokor, Szilárd Aradi, Tamás Bécsi, László Palkovics, Ádám Szabó

    This paper proposes hybrid Neural Network-based Model Predictive Control (NN-MPC) architectures for real-time autonomous vehicle path tracking. To compensate for their individual drawbacks, we integrate supervised learning directly into the optimization process to accelerate solver convergence. Specifically, Multilayer perceptrons are trained to learn a constraint-correction term added to the closedform unconstrained MPC solution, and to provide learning-assisted warm starts for an active-set QP solver by predicting the Lagrange multiplier vector and the active-constraint pattern. The first solution avoids online quadratic programming (QP) and achieves an approximately two orders of magnitude reduction in computation time, at the cost of approximate constraint enforcement in highly transient conditions, whereas the latter two solutions retain the original constrained optimization problem. The learning components are trained on a simplified design model and evaluated in the high-fidelity CarMaker environment to assess robustness under unmodeled dynamics and modeling errors. In CarMaker validation, the classificationbased warm start reduces the number of QP iterations by approximately 30% and computation time by 23% relative to a classical shift-initialization warm start, supporting more predictable real-time operation in representative driving scenarios.


  • Open Access

    Residual Reinforcement Learning Enhanced With Unsuccessful Episode Buffer

    by Peter Farkas, Tamas Becsi, Szilard Aradi

    Reinforcement learning (RL) agents have proven to be highly effective in many control domains; however, despite their generally promising performance, they may still encounter difficulties in certain challenging scenarios. To address this limitation, we propose a two-stage learning framework, where an initial pre-trained policy is extended with a second set of networks, aiming to correct the base agent's knowledge and preserve its behavior. Both modules are trained with the conventional RL loop, while the second learning system is also faced with an increased number of training samples from those problematic scenarios to provide more exploration possibilities across them. We evaluated the effectiveness of the proposed framework on various environments from the robotic and vehicle control domains. The results showed that the final agent demonstrates improved performance and can successfully complete a broader range of scenarios compared to the baseline agents, while also maintaining the control quality.


  • Open Access

    Guided Experience Prioritization for Continuous Reinforcement Learning

    by István Gellért Knáb, Tamás Bécsi, Bálint Kővári

    A primary limiting factor in modern deep learning is the availability of computational resources, a constraint that becomes particularly pronounced in the context of complex reinforcement learning tasks. Hardware performance limitations often result in prolonged training procedures, which are frequently correlated with increased operational costs. In many cases, these limitations place such solutions at a disadvantage, which is why cost reduction remains one of the main priorities. In the field of deep reinforcement learning, a commonly employed method for this purpose is the guided selection of training samples, which emphasizes the most informative state transitions, thereby allowing the agent to maximize its learning from the most critical experiences. However, similarly to the decision-making process, it is also crucial to ensure that the agent has information about which samples are important. Therefore, exploration must be incorporated in this context as well, enabling the agent to acquire the most diverse possible knowledge regarding both the samples themselves and their informational value. The paper introduces a sampling method for continuous-output reinforcement learning agents, which balances, in a problem-independent manner, the necessary exploration and the selection of the most informative samples that enhance learning speed. Due to the presented Upper Confidence Bound-based sampling, both exploration and exploitation are ensured simultaneously, and by tuning their balance appropriately, training can be accelerated. This approach improves the efficiency of the training process, which often constitutes the primary bottleneck.

  • Open Access

    MCTS-Based Policy Improvement for Reinforcement Learning

    by György Csippán, István Péter, Bálint Kővári, Tamás Bécsi

    Curriculum Learning (CL) is a potent field in Machine Learning that provides several excellent techniques for enhancing the performance of the training process given the same data points, regardless of the training method used. In this research, we propose a novel Monte Carlo Tree Search (MCTS)-based technique that enhances model performance, articulating the utilization of MCTS in Curriculum Learning. The proposed approach leverages MCTS to optimize the sequence of batches during the training process. First, we demonstrate the application of our method in Reinforcement Learning, where sparse rewards often diminish convergence and deteriorate performance. By leveraging the strategic planning and exploration capabilities of MCTS, our method systematically identifies and selects trajectories that are more informative and have a higher potential to enhance policy improvement. This MCTS-guided batch optimization focuses the learning process on valuable experiences, accelerating convergence and improving overall performance. We evaluate our approach on standard RL benchmarks, demonstrating that it outperforms conventional batch selection methods regarding learning speed and policy effectiveness. The results highlight the potential of combining MCTS with CL to optimize batch selection, offering a promising direction for future research in efficient Reinforcement Learning.