Training large neural network models requires extensive computational resources, often distributed across several nodes and accelerators. Recent findings suggest that it may be sufficient to only exchange the fast moving components of the gradients, while accumulating momentum locally (Decoupled Momentum, or DeMo). However, DeMo assumes that models fit on a single accelerator. We relax this assumption and introduce FlexDeMo, whereby nodes fully shard model parameters locally between different accelerators, while inter-node communication is reduced by synchronizing only fast-moving components instead of the full gradients -- resulting in a hybrid sharded data parallel training strategy. We further introduce a framework, denoted as DeToNATION, that generalizes DeMo, FlexDeMo, and other popular distributed training schemes such as DiLoCo -- introducing new variations of replication schemes and challenging choices made in DeMo. Our results across language and vision domains show that FlexDeMo attains similar validation loss as hybrid sharded data parallel training employing AdamW and full gradient synchronization, while being substantially faster. FlexDeMo is thus a promising distributed training scheme for the largest machine learning models.

Large-scale datasets are foundational for research and development in natural language processing. However, current approaches face three key challenges: (1) reliance on ambiguously licensed sources restricting use, sharing, and derivative works; (2) static dataset releases that prevent community contributions and diminish longevity; and (3) quality assurance processes restricted to publishing teams rather than leveraging community expertise. To address these limitations, we introduce two contributions: the Dynaword approach and Danish Dynaword. The Dynaword approach is a framework for creating large-scale, open datasets that can be continuously updated through community collaboration. Danish Dynaword is a concrete implementation that validates this approach and demonstrates its potential. Danish Dynaword contains over four times as many tokens as comparable releases, is exclusively openly licensed, and has received multiple contributions across industry and research. The repository includes light-weight tests to ensure data formatting, quality, and documentation, establishing a sustainable framework for ongoing community contributions and dataset evolution.

Contemporary machine learning models, such as language models, are powerful, but come with immense resource requirements both at training and inference time. Quantization aware pre-training with ternary weights (1.58 bits per weight) has shown promising results in decoder-only language models and facilitates memoryefficient inference. However, little is known about how quantization-aware training influences the training dynamics beyond such Transformer-based decoder-only language models. Here, we engage in a bottom-up exploration of quantization-aware training, starting with multi-layer perceptrons and graph neural networks. Then, we explore 1.58-bit training in other transformer-based language models: encoder-only and encoderdecoder models. Our results show that in all of these settings, 1.58-bit training is on par with standard 32/16-bit models, yet we also identify challenges specific to 1.58-bit encoder-decoder models. Our results on decoderonly language models hint at a possible regularization effect introduced by quantization-aware training.

Despite the commercial abundance of UAVs, aerial data acquisition remains challenging, and the existing Asia and North America-centric open-source UAV datasets are small-scale or low-resolution and lack diversity in scene contextuality. Additionally, the color content of the scenes, solar zenith angle, and population density of different geographies influence the data diversity. These factors conjointly render suboptimal aerial-visual perception of the deep neural network (DNN) models trained primarily on the ground view data, including the open-world foundational models. To pave the way for a transformative era of aerial detection, we present Multiview Aerial Visual RECognition or MAVREC, a video dataset where we record synchronized scenes from different perspectives - ground camera and drone-mounted camera. MAVREC consists of around 2.5 hours of industry-standard 2.7K resolution video sequences, more than 0.5 million frames, and 1.1 million annotated bounding boxes. This makes MAVREC the largest ground and aerial view dataset, and the fourth largest among all drone-based datasets across all modalities and tasks. Through our extensive benchmarking on MAVREC, we recognize that augmenting object detectors with ground view images from the corresponding geographical location is a superior pretraining strategy for aerial detection. Building on this strategy, we benchmark MAVREC with a curriculum-based semi-supervised object detection approach that leverages labeled (ground and aerial) and unlabeled (only aerial) images to enhance aerial detection.

In the rural town of Opuwo, Namibia, a team of five young people,, have been working since 2017 on their toy car business. We call them street engineers, because, while they are formally unqualified as engineers, they are informally, or streetwise, qualified because they engineer solutions that meet with their customers' demands, with an inventive twist. Recently, the street engineers joined forces with a university research team to extend their range of products to include robotics. The context was interesting for researchers seeking for innovative practice in engineering education, for several reasons. First, it provides a prime example of a local situation where an increasing number of young talents, despite having limited or irrelevant formal engineering education, globally struggle with designing inventive solutions, engineered of components that are available. Secondly, it poses challenges for devising informal approaches for engineering education, complementing contemporary, typically formal approaches. Thirdly, it calls for engineering education that starts from recognizing young people's hunger for improving their livelihoods through fast-tracked learning. Fourthly, the street engineers aim to develop technologies which, by far, extend the existing state-of-the-art in their context, paving the way for the application and even the design of advanced technologies, especially robotics, at the grassroots, a much promoted political slogan in the Global South. The qualitative analysis of the data sourced in the real-life context of the street engineers identified twelve dimensions of limitations that restrict the work, achievements and outcomes of the street engineers. At the same time, the further interpretation of the limitations indicates a range of opportunities by which design or engineering activities, including entry-level robotics, liberates street engineers from the identified limitations. The context shares the characteristics of places where oppressed people globally live, because of the socio-economic profile of Opuwo and the surrounding Kunene region. Paulo Freire suggests that instead of what he calls the banking model of education, where knowledge is deposited in learners' minds and which, hence, does not transform their living conditions, the oppressed only can be liberated by problem-posing education which aims at identifying challenges and then designing disruptive solutions to them, by those that are limited or bound by their poverty. While obtained within a given context, the results of the investigation can be re-contextualized and used for reforming engineering education in the Global South, to be relevant in the era of the Fourth Industrial Revolution. In an increasingly diverse, global world, robotics invites the much too often marginalized young people to open their minds for and by engineering novel solutions.

With this paper we present the first – to our knowledge – systematic review on how to use Educational Robotics to foster girls’ interest in STEM. This is a research area essential to broadening the participation across the genders in the much-needed STEM workforce, whose size is currently held back by a significant gender disparity. In the review, 13 (quasi-)experimental studies were selected for synthesis, from a total of 1093 results found across multiple search queries applied to six scientific databases. When synthesizing the results and findings from the included studies, four major categories of research interest were identified. On the basis hereof, a list of recommendations, which are readily implementable in most curriculums for both compulsory education and extracurricular activities, was established. The recommendations revolve around: The choice of technology, applied contextualization, approaches to teaching, and gender compositions. In addition, we discuss the current extent of research on the topic, which shows indications of becoming more active in recent years, while likewise discussing the reviews implications for future research directions.

Trust is an important topic in medical human-robot interaction, since patients may be more fragile than other groups of people. This paper investigates the issue of users' trust when interacting with a rehabilitation robot. In the study, we investigate participants' heart rate and perception of safety in a scenario when their arm is led by the rehabilitation robot in two types of exercises at three different velocities. The participants' heart rate are measured during each exercise and the participants are asked how safe they feel after each exercise. The results showed that velocity and type of exercise has no significant influence on the participants' heart rate, but they do have significant influence on how safe they feel. We found that increasing velocity and longer exercises negatively influence participants' perception of safety.

In medical human-robot interactions, trust plays an important role since for patients there may be more at stake than during other kinds of encounters with robots. In the current study, we address issues of trust in the interaction with a prototype of a therapeutic robot, the Universal RoboTrainer, in which the therapist records patient-specific tasks for the patient by means of kinesthetic guidance of the patients arm, which is connected to the robot. We carried out a user study with twelve pairs of participants who collaborate on recording a training program on the robot. We examine a) the degree with which participants identify the situation as uncomfortable or distressing, b) participants' own strategies to mitigate that stress, c) the degree to which the robot is held responsible for the problems occurring and the amount of agency ascribed to it, and d) when usability issues arise, what effect these have on participants' trust. We find signs of distress mostly in contexts with usability issues, as well as many verbal and kinesthetic mitigation strategies intuitively employed by the participants. Recommendations for robots to increase users' trust in kinesthetic interactions include the timely production of verbal cues that continuously confirm that everything is alright as well as increased contingency in the presentation of strategies for recovering from usability issues arising.

In this paper, we demonstrate how a generic industrial robot can be used as a training partner, for upper limb training. The motion path and human/robot interaction of a non-generic upper-arm training robot is transferred to a generic industrial robot arm, and we demonstrate that the robot arm can implement the same type of interaction, but can expand the training regime to include both upper arm and shoulder training. We compare the generic robot to two affordable but custom-built training robots, and outline interesting directions for future work based on these training robots.

We used inspiration from modular robotics to create novel tools for dementia treatment based on activity analyses together with therapists and elderly in an Italian home care. This paper presents the technological development of such tools. In general, the tools are becoming part of a "multi-sensory room", i.e. a space augmented by innovative technologies, that can be configured for different therapeutic activities and needs and that leverages sensory stimulation. The development of the tools is inspired by modular robotics in order to allow for space re-configurability and adaptivity, which should support customized therapeutic interventions, and hopefully involve dementia affected users in the interaction with the solutions

Meaningful learning of information and communication technology (ICT) requires students to understand ICT concepts in their own cultural or societal context. In this study, we hypothesized that grounding ICT education in the local context and using concretizing tools, like programming visualization and intelligent building blocks (I-BLOCKs), would benefit students. We investigated this hypothesis by first developing an introductory ICT course in programming, which applied these strategies, and then by analyzing the results of the course. The course was held at Tumaini University, Iringa University College, Tanzania, in the fall of 2004. Altogether 27 teacher-trainees participated in the course. The outcomes of the course showed that the students understood the main concepts that were taught. They were also able to apply the methods of creative problem solving that were taught in the course. It was also found that students' algorithmically-oriented understanding was not at the desired level and that the course, as well as the future curriculum, needs further development in this particular area.

We developed spiking neural network control for a modular robotic system. The modular robotic system can be easily assembled by a user who is allowed to make overall behaviors by assembling the physical structure made up of a number of modules. The control of each module (building block) is implemented as a spiking neuron and action potentials are sent through the communication channels of the building blocks. We show how to make a mobile robot with these spiking neural building blocks. Hebbian learning is then applied to the spiking neuron building blocks in order to allow the mobile robot to adapt to changing environmental conditions. Collected data shows the learning process, sensor adaptation and performance on a simple task for the mobile robot made out of spiking neural building blocks.