Early Career Researchers

  •
    Nícolas Francisco Figueroa Mosquera

    RoboticsLab, NFM Robotics, Peru

    A pioneer and ecosystem architect in robotics, with a career dedicated to connecting cutting-edge research, industrial application, and technology policy. My mission is to accelerate technology adopti...

    A pioneer and ecosystem architect in robotics, with a career dedicated to connecting cutting-edge research, industrial application, and technology policy. My mission is to accelerate technology adoption and cultivate the next generation of leaders in Latin America, operating from a unique perspective that integrates business, global standards (IEEE), and government advisory services.

    Talk Title: Fast Autolearning for Multimodal Walking in Humanoid Robots with Variability of Experience

    Abstract: Achieving stable, human-like walking in complex and unpredictable environments remains a primary challenge in humanoid robotics. This presentation introduces a novel reinforcement learning framework that significantly enhances a robot's ability to adapt and maintain balance in real-time. Our approach, termed Variability of Experience (VE), integrates two key concepts: a criterion for detecting changes in tasks like locomotion or manipulation, and an efficient method for accumulating and storing these learned experiences. By combining advanced techniques like Proximal Policy Optimization (PPO) and Model-Agnostic Meta-Learning (MAML), our method enables the robot to learn rapidly from new situations. We validated this framework through extensive simulations and real-world experiments on the HRP-4 humanoid robot. The results demonstrate a statistically significant reduction in key stability errors (such as ZMP and DCM) when navigating uneven surfaces and handling objects, showcasing a more robust and anticipatory response compared to baseline methods. This work represents a key step towards more agile and reliable humanoid robots capable of operating in complex, real-world scenarios.

  •
    Rodrigo Pérez Dattari

    KTH, Sweden

    I'm Rodrigo Pérez-Dattari, a researcher interested in every aspect of decision-making for intelligent agents. I'm currently working as a postdoctoral researcher at the KTH in Sweden. Previously, I com...

    I'm Rodrigo Pérez-Dattari, a researcher interested in every aspect of decision-making for intelligent agents. I'm currently working as a postdoctoral researcher at the KTH in Sweden. Previously, I completed my doctoral studies in Delft, Netherlands, in 2023 (defended in September 2024). My work has primarily focused on developing methodologies to enhance the learning of robot policies, with the aim of achieving greater flexibility, efficiency, and reliability.

    Talk Title: Imitation Learning through the Lens of Artificial Dynamical Systems with Robotic Priors

    Abstract: Despite the significant research and development conducted in robotics in the last few decades, its full potential in practical real-world applications remains largely untapped. Currently, most real-world robotic systems employ simplistic methods, such as using pre-recorded waypoints to define a robot's trajectory. Although such approaches have been successful in highly structured and predictable environments, their effectiveness quickly diminishes in less controlled settings, e.g., the agro-food industry. One major obstacle limiting the advancement of more sophisticated solutions is the complexity and cost associated with their development and implementation. Hence, to effectively create solutions that can be widely adopted for solving the current real-world robotic challenges, we require robots that are easily programmable and capable of seamlessly performing tasks in unstructured and variable environments. Moreover, anyone, regardless of their technical background, should be able to achieve this. Then, the cost of creating complex systems would be reduced significantly, also enabling the incorporation of robots in settings where access to robotics expertise is limited, such as in small to medium-sized businesses or household environments. To achieve this objective, Imitation Learning (IL) is an excellent candidate. As a data-driven approach, IL facilitates the transfer of human expertise to robots through intuitive human-robot interactions. This significantly reduces the complexity of programming robotic behaviors, which would otherwise require intricate rule sets or the meticulous design of reward/cost functions. Despite significant progress in IL, one fundamental issue persists: when deployed, our robots are likely to encounter states not represented in the training data, resulting in poor generalization and, therefore, failure. The basic nature of this issue is deceptively simple, which leads to the assumption that simply collecting additional training data should solve this problem. Unfortunately, experience has shown that merely collecting more data is far from straightforward. This challenge arises from several issues, including 1) errors that compound, causing a mismatch between training and testing distributions, known as covariate shift, 2) tasks that involve numerous unlikely situations, leading to long-tailed state distributions, and 3) practical constraints that limit the resources available for data collection. In this presentation, I will focus on one of the approaches I have followed to tackle these challenges: incorporating inductive biases through tools from control theory and differential geometry. With this, we aim to introduce prior knowledge into a robot’s behavior. The objective is to identify and integrate aspects of the robot’s behavior that remain consistent throughout a task and embed them as intrinsic elements of its policy. As a result, even when encountering unfamiliar scenarios, a robot endowed with such behaviors is more likely to execute reasonable actions. This greatly increases the generalization capabilities of our robots and the data efficiency of our learning algorithms.

  •
    Angel Antonio Ayala Maldonado

    Universidade de Pernambuco, Brasil

    PhD Student at Universidade de Pernambuco, Brasil ...

    PhD Student at Universidade de Pernambuco, Brasil

    Talk Title: Decoupled sensorimotor self-predictive representation for object-goal autonomous navigation with a quadcopter

    Abstract: Novel solutions that deploy autonomous unmanned aerial vehicles (UAVs) rely on intelligent autonomous navigation systems (ANS). According to systematic literature review (SLR) findings, an ANS deals with vehicle travel from one point to another by solving control, planning, mapping, and perception tasks. The object-goal ANS becomes challenging since it extends the four tasks in pursuit of a target object. A second finding highlights that most methods nowadays utilize autopilot boards for control, graph-based methods for planning, SLAM or occupancy maps for mapping, and handcrafted features for perception. However, one of the major findings from the SLR was the low reuse of perception model information from planning, control, and mapping tasks. Therefore, a research question and its corresponding hypothesis arose to define the research objectives in the area of probabilistic robotics. Probabilistic methods have been widely used in robotics and have evolved into a field known as probabilistic robotics. Modern approaches are so named because they are learning-based methods, mostly neural network models that use the backpropagation algorithm for parameter optimization. More recently, reinforcement learning (RL) approaches have emerged as a promising option for decision-making problems, yielding outstanding outcomes. However, its main challenge is the inefficiency of the data sample in achieving a good policy, which worsens with higher-dimensional observations. Advances in deep learning models have demonstrated their ability to acquire rich features through representation learning approaches. The features from the representation model optimized in a self-supervised manner have been successfully applied to other downstream tasks. Hence, the research hypothesizes that state representation learning (SRL) approaches can efficiently solve the object-goal autonomous navigation problem using a quadcopter. In object-goal navigation settings, another challenge related to target recognition is also present. Object-goal navigation is a topic that has been extensively researched in embodied methods for indoor settings. Most UAV-related approaches use relative or absolute coordinates to move from an initial position to a predefined location, rather than directly finding the target. Therefore, the research scope also addressed the formalization of the object-goal navigation problem with an undetermined target location as a Markov Decision Process. Hence, the main challenges addressed in this research are related to data sample efficiency in solving a 3D object-goal navigation (OGN) problem. Conducted research, analyzed the interplay of different SRL methods for the perception task with a model-free RL algorithm for the planning task, and the UAV inner-controller for the control task. The proposed object-goal ANS was named Chemamuy, inspired by the Mapuche people, meaning 'Someone is moving towards that.' The main contribution of this research was the development of the perception module, featuring a novel self-predictive model named AmelPred, in addition to the decoupled AdPuEncoder. The main proposal, as well as different combinations of RL and SRL, were all evaluated under an OGN setting. The OGN scenario comprises a UAV as an IoT device that must autonomously fly towards a goal pose near the target location. A total of 24 different target locations in a 3D space were used to evaluate the navigation performance of each method. Experiments comparing the combination of DQN, SAC, and TD3 algorithms for the planning task, incorporating features from reconstruction or prediction-based techniques, were addressed alongside this research. Empirical results demonstrated that AmelPred was the best-performing SRL model combined with any model-free RL algorithms. The use of a stochastic encoder function achieved even better results for actor-critic models. Additionally, the proposed extension using AdPuEncoder encoders with AmelPred was also validated to test the decoupled sensor improvement hypothesis. AmelPredSto-APE was the most data-sample efficient SRL approach for SAC and TD3 algorithms in OGN problems. The research's findings aim to contribute to the knowledge of quadcopter autonomous navigation by studying the impact of SRL approaches in object-goal settings. The obtained results support the proposed hypothesis by showing the improvement of RL algorithms by using SRL techniques in solving the object-goal navigation problem. Future works can address a fusion method for the correct reuse of each proprioceptive and exteroceptive encoder, as well as reusing pretrained deep learning models' knowledge to acquire rich features from images.

  •
    Javier Andre Borquez Berenguer

    USACH, Chile

    Visiting Scholar at Stanford University...

    Visiting Scholar at Stanford University

    Talk Title: Safety Analysis for Real World Autonomous Systems: From High Dimensional and Hybrid Systems to Online Adaptation

    Abstract: As autonomous systems become increasingly integrated into society, ensuring their ability to operate safely in complex, uncertain, and dynamic environments is essential. This dissertation develops a suite of methods that equip such systems with formal safety guarantees while preserving high-performance behavior. Grounded in Hamilton-Jacobi (HJ) reachability, the proposed approaches provide principled frameworks for computing safe sets and designing safety-preserving interventions. To bridge the gap between offline safety analysis and real-world deployment, we introduce real-time safety filters that minimally modify nominal control policies to enforce safety or liveness constraints. These tools are further integrated into sampling-based optimal control frameworks to enable goal-directed, safety-critical planning in real time. The dissertation also extends reachability-based safety analysis to accommodate uncertain and hybrid system dynamics, proposing mechanisms for safe control under model uncertainty and for task completion in systems with both continuous and discrete transitions. Finally, we develop adaptive reachability methods that allow safety guarantees to evolve online in response to changes in environment or task parameters. Together, these contributions advance the theory and practice of safe autonomy, enabling robust, goal-driven control in the face of uncertainty, hybrid dynamics, and real-time variability.

  •
    Jhon Ivan Pilataxi Pilataxi

    Universidad de Chile, Chile

    PhD student University of Chile...

    PhD student University of Chile

    Talk Title: Neural Architecture Classifier for Improving Search in Neuroevolution Applied to Pattern Recognition

    Abstract: The results of the PhD will be presented. The thesis proposes a novel Genetic Algorithm (GA) that integrates a Neural Architecture Classifier (NAC) to reduce the computational cost of Neuroevolution. The NAC classifies each individual in the current population, representing Convolutional Neural Networks (CNNs), as either high performing or low-performing. A performance predictor model is used to estimate the fitness of CNNsclassifiedaslow-performing, while high-performing CNNs are trained from scratch to obtain a more accurate fitness. This approach avoids training the entire population, thereby reducing the computational cost. Additionally, a search space with minimal restrictions is introduced, allowing the representation of architectures with different depths, widths, or shapes. The NAC is trained using architectures evaluated in previous generations, using architecture-encoded representations as a feature vector. Experimental results demonstrate that incorporating the NAC into the GA not only reduces search time by up to 44% but also enhances search space exploration, finding better architectures than experiments without NAC. The proposed method was applied to two pattern recognition tasks: image classification and low-resolution face recognition, using publicly available datasets, achieving results that outperform the state-of-the-art on MNIST-Variations (by 4% on average), QMUL-TinyFace (by 2.88% on rank-1 recognition rate), and QMUL-SurvFace (by 12.6% on AUC metric).