Accurate human trajectory prediction is one of the most crucial tasks for autonomous driving, ensuring its safety. Yet, existing models often fail to fully leverage the visual cues that humans subconsciously communicate when navigating the space. In this work, we study the benefits of predicting human trajectories using human body poses instead of solely their Cartesian space locations in time. We propose 'Social-pose', an attention-based pose encoder that effectively captures the poses of all humans in a scene and their social relations. Our method can be integrated into various trajectory prediction architectures. We have conducted extensive experiments on state-of-the-art models (based on LSTM, GAN, MLP, and Transformer), and showed improvements over all of them on synthetic (Joint Track Auto) and real (Human3.6M, Pedestrians and Cyclists in Road Traffic, and JRDB) datasets. We also explored the advantages of using 2D versus 3D poses, as well as the effect of noisy poses and the application of our pose-based predictor in robot navigation scenarios.

Tracking hyperpolarized (HP) ¹³C labeling from either [1-¹³C] pyruvate or [1-¹³C] lactate is a useful tool to assess intermediary metabolism in vivo, which has already been translated from preclinical to clinical research. HP [1-¹³C] pyruvate and [1-¹³C] lactate provide complementary views on the same metabolic pathway, and both have been tested as potential neuroprotective agents in the context of acute brain injuries, with more convincing evidence for a beneficial effect of lactate. Our aim here was to investigate and compare HP [1-¹³C] pyruvate and [1-¹³C] lactate performance as metabolic contrast agents in the brains of healthy mice and mice subjected to middle cerebral artery occlusion, a model of ischemic stroke. We analyzed the metabolite ratios and quantified the real-time apparent kinetic rates of their cerebral metabolism. We found that the cerebral metabolism of both HP [1-¹³C] pyruvate and HP [1-¹³C] lactate showed significant alterations after transient cerebral ischemia in mice, reflecting the damage as well as the metabolic reprogramming set in motion to meet the energetic demands in the acute phase of stroke. There was a significant decrease in metabolite ratios (cLPR, cAPR for pyruvate bolus and cPLR, cALR for lactate bolus) and kinetic rates (ckPL for pyruvate bolus and ckLP for lactate bolus). These values progressively decreased from sham to 1 h and 2 h after reperfusion measurements. Overall, while pyruvate is better established as an imaging probe, and lactate appears advantageous on the therapeutic side, both bring information to interrogate brain metabolism in physiological and pathophysiological conditions in real time. This study prepares the ground for further investigation to fully exploit the potential of HP metabolic contrasts for stroke theranostics.

EPFL has processed data on scholarly publications published between 2021 and 2024 and written by affiliated researchers. The bibliographic data were taken from EPFL’s institutional repository, Infoscience (https://infoscience.epfl.ch/), and enriched with metadata from OpenAlex (https://openalex.org/) to classify publications into four models (gold, hybrid, green and closed) in response to the Swiss Open Access Monitor’s annual survey (Repository Monitor – https://oamonitor.ch/charts-data/repository-monitor/). The datasets comprise four publication types (journal articles, books, book sections and conference papers). Annual data sets are provided as individual CSV files in this record.

Raw data accompanying the paper.

Metal-arene complexes have recently attracted an increasing interest in f-element chemistry, but the functionalization of arenes mediated by uranium-arene interactions is limited to a single example. Here, we report a new uranium-biphenylene complex supported by a bulky rigid trianionic NNN-pincer ligand in which the uranium-arene interaction is able to promote C−C bond cleavage and functionalization with CO under mild conditions to yield a U-bound 9-fluorenone. Reduction of the U(IV)-pincer complex [NNN-U(THF)Cl2K(THF)3]2 (1) with KC8, in the presence of biphenylene, results in the terminal arene complex [NNN-U(THF)(biphenylene)][K(THF)5] (3). DFT studies of 3 indicate the presence of two unpaired electrons located at the uranium center, in line with a U(IV) and a biphenylene dianion. Complex 3 undergoes Caryl−Caryl bond cleavage of the biphenylene ligand, affording [NNN-U(THF)(2,2'-biphenyl)][K(THF)2] (4). DFT studies indicated that, due to the interaction between the biphenylene dianion and the uranium, a concerted ring opening reaction can occur on the strained four members ring to yield 4 while the uranium center retains a +IV oxidation state. Complex 4 undergoes facile CO insertion into the U−Caryl bond, followed by the Caryl−Ccarbonyl bond formation, yielding [NNN-U(THF)2(fluorenone)][K(THF)4] (5). This work demonstrates the potentials of uranium-arene interactions to promote arene activation and functionalization

Deployable structures that transform from a planar assembly-friendly compact state to an expansive freeform surface state have diverse applications in robotics, medical devices, temporary installations, and architecture. Umbrella Meshes are a new class of volumetric deployable structures with extensive shape expression capabilities compared to existing plane-to-surface deployables. They are modular, made of Umbrella cells consisting of identical rigid plates and rotational joints connected by elastic beams of varying heights. Deployment is actuated by pushing the cells orthogonal to the plane, rotating the elastic beams from vertical to horizontal configurations, thus redistributing material from out of the plane into it. In contrast to rigid scissor mechanisms, the beams deform elastically, making the deployed equilibrium bending-active. Assembled in a stress-free planar configuration, an Umbrella Mesh can be programmed to deploy to a desired target shape by virtue of the optimized heights of the constituent cells. The rich design space facilitates programming a large range of target shapes, controlling the structural stiffness, and encoding extrinsic curvature.

This thesis contributes a comprehensive computational framework for the design and optimization of Umbrella Meshes. To facilitate design exploration of the deployed structure, we develop a physics-based simulation modeling the deployment process under actuation forces. We abstract the deployment transformation of an umbrella mesh using conformal geometry, providing intuitive design initializations for a specific target surface. Our inverse design algorithm leverages the simulation pipeline and numerical optimization to iteratively refine a design to approximate a target surface while minimizing the elastic energy and actuation forces involved. We build optimized physical prototypes through digital fabrication and validate our computational pipeline.

The inverse design framework exemplifies a design-driven approach to fabricating optimized physical structures. The latter half of this thesis focuses on fabrication-driven design. We develop a computational framework to rationalize bending-active structures into a sparse kit of parts, allowing cost-effective fabrication. Our method can either find an optimal kit of parts for multiple input designs or rationalize existing designs to use a pre-fabricated kit of parts. To tackle the non-trivial coupling of components in bending-active systems, we propose a relaxed continuous formulation of the combinatorial problem of grouping components to a sparse part set, allowing us to incorporate physics-based simulation that tracks multiple bending-active equilibria. We demonstrate our approach on Umbrella Meshes, C-shells, and orthogonal gridshells.

The thesis culminates with Reconfigurable Umbrella Meshes (RUMs) consisting of identical reconfigurable cells. Each reconfigurable cell can assume the form of a continuous range of parts, thus combining the benefits of pre-fabrication and precisely inverse-designed heights. Assembled from these identical mass-producible cells, the same RUM can deploy into several shapes over multiple deployment cycles. Our inverse design enables precise reconfiguration of the compact state and opens up multiple research avenues for high-fidelity shape morphing control with applications in soft robotics and sustainable architecture.

To meet the Paris Agreement target of limiting global warming to 1.5°C, the International Energy Agency stresses the urgent need for rapid and transformative actions in all sectors. In 2022, transportation accounted for about 25% of global CO2 emissions, making decarbonization a critical priority. Among transportation modes, rail is the least carbon-intensive and is expanding significantly worldwide. High-speed rail, in particular, provides a viable alternative to short- and medium-haul flights, helping reduce aviation-related emissions. Innovative, sustainable transportation technologies are gaining interest as complementary solutions to rail expansion. They diversify land transportation and help reduce aviation-related emissions. Governments, especially the European Commission, have shown renewed commitment to advancing high-speed systems. This political interest has brought back interest in established but underutilized technologies like Hyperloop and maglev trains. The propulsion of maglev and Hyperloop systems is typically achieved using Linear Electrical Machines (LEMs), which are types of electrical machines that produce linear motion by generating a direct and contactless thrust force along a straight-line path. Among various types of LEMs, Linear Induction Motors (LIMs) stand out as promising candidates for maglev propulsion. LIMs correspond to the linear counterpart of conventional Rotating Induction Motors (RIMs) and may offer significant advantages compared to other LEMs, such as simpler construction, lower cost, and scalability. However, LIMs have traditionally been restricted to low-speed, short-haul applications due to their lower efficiency and gravimetric force densities than other LEMs. This thesis explores the potential of LIMs for medium- to high-speed maglev ground transportation systems, focusing on the integration of Propulsion and Levitation (PL) or Propulsion and Guidance (PG) functionalities into a single motor for an all-in-one maglev system. The core of the thesis is the development of a highly accurate and computationally efficient analytical model of LIMs that allows for the calculation of motor electromagnetic fields, forces, and efficiency. The proposed analytical model has been validated through comparisons with Finite Element Analysis (FEA) simulations and measurements from a custom-made experimental platform, demonstrating excellent accuracy and superior computational efficiency compared to FEA models. The proposed analytical model is utilized in the thesis to enhance the performance of Single-Sided LIMs (SLIMs), focusing particularly on increasing their gravimetric force densities, and to demonstrate the potential of SLIMs for a MHS maglev system with PL or PG functionalities integrated into the same motor. The thesis also proposes an optimization framework for the design of SLIMs, in which the developed analytical model and the performance enhancement techniques mentioned above have been combined into a multiobjective optimization problem. The objective is to maximize the Levitation-to-Weight Ratio (LWR) and the efficiency of SLIMs for a reduced-scale Hyperloop prototype operated at the EPFL Hyperloop test infrastructure. Finally, a control strategy is proposed to achieve a decoupled, simultaneous, and electromagnetic drag-less control of PL in SLIMs, thereby unlocking their potential to combine these functionalities into a single motor.

Contemporary AI faces important limitations that remain unresolved despite significant successes. Chief among these are the inability to acquire new knowledge without destroying old, the incomprehensible and nonengineerable nature of internal representations, and the difficulty of integrating learned models with symbolic reasoning. As a result, essential capabilitiesâ continual open-world adaptation, recursive improvability, and transparent, human-controllable autonomyâ remain out of reach. This thesis argues that overcoming these challenges requires a rethinking of the foundations of artificial learning. Toward that end, it proposes foundations of a new paradigm for learning in AI grounded in principles of adaptation revealed by evolutionary developmental biology.

The first part lays the conceptual groundwork by drawing parallels between AI and evolutionary theory. It critiques foundations of the current mainstream AI approachesâ fixed architectures, distributed representations, statistical optimizationâ showing how aforementioned limitations follow directly from these. The discussion then turns to the Modern Synthesis of evolution, which, though powerful, fails on its own to explain accelerating evolutionary change and the structural properties of biological organization. These gaps have been addressed in the "extended evolutionary synthesis," notably through insights from developmental biology. By aligning the explanatory limitations of the Modern Synthesis with capability limitations in AI, this section argues that principles discovered by evolutionary developmental biology offer a promising foundation for reimagining how learning systems are built.

The second part operationalizes one such principleâ conditional regulationâ within the current learning paradigm. It introduces the Directed Adaptation Network (DIRAN), a method that incrementally constructs a continuously-parameterized structure learning via gradient signal without resorting to heavy overparameterization. Instead of tuning a fixed architecture, DIRAN generates regulatory connections through a developmental process driven by conflicting learning pressures, and guarantees convergence to good solutions with minimal complexity.

The final part presents the thesisâ s main technical contribution: the foundations and early design on a new class of learning systems called \textit{variation and selection (varsel) networks}. In addition to incorporating the principle of regulability, these systems reconceive learning as the generation of structured explanations through local component-level variation and selection instead of iterative optimization. As a non-gradient-based method, this approach generates weakly-linked, topologically organized representations that support continual learning without any constraining assumptions, decomposability and interpretability, and natural integrability with symbolic processes. Demonstrative experiments on those capabilitiesâ including modeling the dynamics of a simple environment, behavior integration, and continual learning in a visual spaceâ demonstrate these properties and chart a path toward more adaptive, transparent, and controllable AI systems. Rooted in the first principles of biology, this paradigm offers a unified foundation for a system that is not constrained by the core limitations of current methods.

Semiconductor innovation is shifting from traditional dimensional scaling toward novel architectures and materials as lithography nears its physical limits. FinFETs and GAAFETs have marked key milestones, with scaling now progressing through 3D stacking. Meanwhile, 2D materials like MoS2 offer potential for energy-efficient transistors with increased density. However, beyond-silicon technologies face challenges in industrial integration, primarily due to the lack of CMOS-compatible and scalable fabrication processes for high-yield nanoelectronics and their low electron mobilities (20-30 cm2/V.s), which remain far below the industry target (>100 cm2/V.s).

In this thesis, we explore the third dimension in device architecture through grayscale nanopatterning, enabling engineered surface topographies for the strain engineering of 2D materials. For grayscale nanopatterning, we use thermal scanning probe lithography, which stands out due to its sub-1 nm depth control, but also has inherent challenges, including geometry-dependent spatial resolution, limited patterning depth and aspect ratio, low throughput due to tip scanning, and substrate constraints arising from electrostatic actuation. To expand its potential, we demonstrate: (1) high-aspect-ratio metal tip integration for deeper patterning; (2) grayscale pattern transfer and depth amplification into thin film dielectrics using gentle plasma etching; and (3) scalable replication via nanoimprint lithography with smooth pattern transfer onto various substrates.

We apply these grayscale nanosurfaces to strain-engineer monolayer MoS2 for enhanced electron mobility through two fundamentally distinct approaches: (1) transfer-based strain and (2) transfer-free strained growth. In the transfer-based approach, we achieved electron mobilities up to 185 cm2/V.s (~8x improvement with ~1% tensile strain). However, to eliminate interface issues and contamination caused by polymer-assisted 2D material transfer, we develop a new industry-compatible transfer-free approach: introducing strain in 2D materials during their growth on grayscale-patterned surfaces instead of flat substrates, where the grayscale-thin-film/substrate stack is engineered based on thermal expansion mismatches. This method enables precise control of both strain levels (0-0.5% tensile) through aspect ratio modulation and strain orientations (uniaxial and multiaxial) through topography design, and results in a reproducible and scalable technique that creates contamination-free and air-trap-free semiconductor/dielectric interfaces essential for advanced electronics.

The demonstrated fabrication and strain-engineering techniques present scalable and industry-compatible approaches for realizing high-performance 2D material devices with CMOS integration potential. Moreover, scalable strain engineering in 2D semiconductors holds the potential for both logic device scaling and application innovations in the rapidly growing field of see-through/transparent electronics and optics, capabilities beyond the reach of Si-based technologies.

This thesis explores the intersection of physiological modelling and computational techniques in advancing Automatic Speech Recognition (ASR) systems. Contemporary ASR, often driven by attention models and self-supervised learning, has achieved remarkable accuracy, but remains decoupled from more recent physiological principles. In the meantime, significant progress has been made in understanding the function of the cochlea, the auditory system's sensory organ. Originally viewed as a passive filter bank, the cochlea is now understood to function as an active amplifier, well modelled by a Hopf oscillator.

The goal of this thesis is to investigate how such advances in physiological understanding can be studied in the context of such state of the art ASR techniques. To this end, the thesis is organised as two interacting threads.

In a first thread, we investigate modularity, which proposes strategies to integrate and combine different types of machine learning models, using different experts, or combine new frontend models with pretrained large transformer models. In a preliminary study, we show that modularity can be used to optimise an ASR model for different types of environmental noise.

In a second thread, we utilise modularity to investigate how to incorporate improved cochlear understanding into ASR systems, creating a two-way bridge where insights from computational approaches inform auditory physiology. After studying established techniques such as CARFAC and SincNet, we investigate trainable filter banks within a convolutional neural network (CNN) structure to determine key hyperparameters for ASR performance. This study also highlights interesting insights filters tend to learn when able to train in an ASR context.

Finally, we combine the threads by embedding a Hopf-based cochlear model within an ASR system, informed by the learned filter bank parameters. We show that the Hopf mechanism demonstrates the expected cube root compression and gain control. Moreover, a larger feedback loop, modelling the olivocochlear efferent path further enhances the overall performance. The resulting system, offers valuable insights for future interdisciplinary studies between ASR and physiological auditory models.