Evolutionary robotics commonly employs quality-diversity algorithms to discover diverse, high-performing behaviours. Meanwhile, central pattern generators provide a compact neural dynamical model for robotic locomotion control. Their differential-equation-based formulation induces smooth, rhythmic actions, while also reducing the dimensionality of the search space. However, the widely used Kuramoto-inspired central pattern generators exhibit nonlinear dynamics, for which convergence depends sensitively on network connectivity and coupling parameters. When convergence is slow or unstable, controller evaluations become noisy, which substantially degrades quality-diversity performance. This work introduces an analytical criterion to assess and guarantee the convergence properties of Kuramoto-based central pattern generators prior to optimisation or simulation. The method yields principled guidelines for selecting connectivity structures and coupling strengths based on the eigenvalues of the Laplacian of the connectivity graph. Experiments on a simulated brittle star robot demonstrate that adhering to this criterion, significantly improves the efficiency and discovery of coordinated gaits. The criterion (a) predicts stable convergence from network structure alone, (b) enhances both the breadth and quality of exploration, and (c) scales naturally to morphologies of arbitrary complexity.

Self-mixing interferometry (SMI) has been lauded for its sensitivity in detecting microvibrations, while requiring no physical contact with its target. In robotics, microvibrations have traditionally been interpreted as a marker for object slip, and recently as a salient indicator of extrinsic contact. We present the first-ever robotic fingertip making use of SMI for slip and extrinsic contact sensing. The design is validated through measurement of controlled vibration sources, both before and after encasing the readout circuit in its fingertip package. Then, the SMI fingertip is compared to acoustic sensing through four experiments. The results are distilled into a technology decision map. SMI was found to be more sensitive to subtle slip events and significantly more resilient against ambient noise. We conclude that the integration of SMI in robotic fingertips offers a new, promising branch of tactile sensing in robotics. Design and data files are available at https://github.com/RemkoPr/icra2025-SMI-tactile-sensing.

Foundation models, such as ULIP-2 (Xue et al., 2023) recently projected forward the field of 3D deep learning. These models are trained with significantly more data and show superior representation learning capacity in many downstream tasks like 3D shape classification and few-shot part segmentation. A particular characteristic of the recent 3D foundation models is that they are typically multi-modal, and involve image (2D) as well as caption (text) branches. This leads to an intricate interplay that benefits all modalities. At the same time, the nature of the 3D encoders alone, involved in these foundation models is not well-understood. Specifically, there is little analysis on the utility of both pre-trained 3D features provided by these models, or their capacity to adapt to new downstream 3D data. Furthermore, existing studies typically focus on label-oriented downstream tasks, such as shape classification, and ignore other critical applications, such as 3D content-based object retrieval. In this paper, we fill this gap and show, for the first time, how 3D foundation models can be leveraged for strong 3D-to-3D retrieval performance on seven different datasets, on par with state-of-the-art view-based architectures. We evaluate both the pre-trained foundation models, as well as their fine-tuned versions using downstream data. We compare supervised fine-tuning using classification labels against two self-supervised label-free fine-tuning methods. Importantly, we introduce and describe a methodology for fine-tuning, as we found this to be crucial to make transfer learning from 3D foundation models work in a stable manner.

We present KeyCLD, a framework to learn Lagrangian dynamics from images. Learned keypoints represent semantic landmarks in images and can directly represent state dynamics. We show that interpreting this state as Cartesian coordinates, coupled with explicit holonomic constraints, allows expressing the dynamics with a constrained Lagrangian. KeyCLD is trained unsupervised end-to-end on sequences of images. Our method explicitly models the mass matrix, potential energy and the input matrix, thus allowing energy based control. We demonstrate learning of Lagrangian dynamics from images on the cl_bnmsqnk pendulum, cartpole and acrobot environments. KeyCLD can be learned on these systems, whether they are unactuated, underactuated or fully actuated. Trained models are able to produce long-term video predictions, showing that the dynamics are accurately learned. We compare with Lag-VAE, Lag-caVAE and HGN, and investigate the benefit of the Lagrangian prior and the constraint function. KeyCLD achieves the highest valid prediction time on all benchmarks. Additionally, a very straightforward energy shaping controller is successfully applied on the fully actuated systems.

This paper explores the potential of self-supervised learning as an alternative to supervised learning in the context of geometry-based 3D object retrieval. With the ongoing digitalization of many industries, an exponentially increasing number of 3D objects are processed by retrieval systems. In order to support new shapes, modern deep learning-based retrieval systems require retraining. The dominant paradigm for optimizing neural networks in this field is supervised classification training. Supervised learning requires time-consuming and expensive data annotation. Moreover, training neural networks for classification introduces a bias towards the classes in the training data, which is undesirable for retrieval systems encountering unseen object types in the wild. Through extensive experiments, we make a direct comparison between supervised and self-supervised learning on four datasets from three different domains (household, manufacturing and medical). For object classes seen during training, self-supervised and supervised learning are competitive. For unseen classes, self-supervised learning outperforms supervised learning in many cases. We conclude that self-supervised learning provides a powerful tool for circumventing labeling costs and providing more robust retrieval systems.

In many countries, computer programming is becoming an integral part of the secondary school curriculum. However, many teachers, especially in the first years of Flemish secondary school, have limited experience with teaching programming. To improve their knowledge about programming, many different types of professional development programs have been proposed. Nevertheless, these programs mostly focus on technical skills and less on pedagogical skills. One aspect that is often overlooked in these programs is how teachers can assess code. To get insight into what teachers currently value when assessing code, we designed an experiment that analyzes the different aspects teachers consider during the assessment of code. During the experiment, the teachers (N=13) assess a set of programs from five different fictional learners. After the assessment, they participated in a semi-structured interview, giving us insight into the assessment process. We evaluated the transcripts of the interviews using deductive thematic analysis using a coding schema defining the different aspects of code that can be assessed. Additionally, we linked the assessment strategies of teachers to their teaching experience. Our results indicate that many teachers are unaware of the different concepts that can be part of the assessment of code, which might lead to inaccurate or invalid feedback. Moreover, although our experimental group was too small to draw hard conclusions about the inter-case results, our results indicate that the number of concepts considered by teachers seems to increase with experience. These results provide an initial insight into the code assessment practices of teachers and reveals interesting pathways for future research into the assessment of code.

Advances in legged robotics are strongly rooted in animal observations. A clear illustration of this claim is the generalization of Central Pattern Generators (CPG), first identified in the cat spinal cord, to generate cyclic motion in robotic locomotion. Despite a global endorsement of this model, physiological and functional experiments in mammals have also indicated the presence of descending signals from the cerebellum, and reflex feedback from the lower limb sensory cells, that closely interact with CPGs. To this day, these interactions are not fully understood. In some studies, it was demonstrated that pure reflex-based locomotion in the absence of oscillatory signals could be achieved in realistic musculoskeletal simulation models or small compliant quadruped robots. At the same time, biological evidence has attested the functional role of the cerebellum for predictive control of balance and stance within mammals. In this paper, we promote both approaches and successfully apply reflex-based dynamic locomotion, coupled with a balance and gravity compensation mechanism, on the state-of-art HyQ robot. We discuss the importance of this stability module to ensure a correct foot lift-off and maintain a reliable gait. The robotic platform is further used to test two different architectural hypotheses inspired by the cerebellum. An analysis of experimental results demonstrates that the most biologically plausible alternative also leads to better results for robust locomotion.

The Progra-MEER professional development workshop is a one year program organized collaboratively by the computer science departments of three Flemish universities. It aims to improve the computer science knowledge of in service teachers in a physical computing context. Since Flemish schools are starting to implement STEM in their schools, the program links computer science to STEM and project based learning. This paper gives a description of the design and implementation of the program while providing an analysis of its strengths and weaknesses. We show that the program leads to the successful implementation of different physical computing projects. However, it needs to further support the practical project implementations while spending more attention on assessment and context definition. Additionally, the program has to invest more effort in creating a sustainable community of practice so knowledge and experiences can still be shared even after the program has finished.

Incorporating the body dynamics of compliant robots into their controller architectures can drastically reduce the complexity of locomotion control. An extreme version of this embodied control principle was demonstrated in highly compliant tensegrity robots, for which stable gait generation was achieved by using only optimized linear feedback from the robot's sensors to its actuators. The morphology of quadrupedal robots has previously been used for sensing and for control of a compliant spine, but never for gait generation. In this paper, we successfully apply embodied control to the compliant, quadrupedal Oncilla robot. As initial experiments indicated that mere linear feedback does not suffice, we explore the minimal requirements for robust gait generation in terms of memory and nonlinear complexity. Our results show that a memoryless feedback controller can generate a stable trot by learning the desired nonlinear relation between the input and the output signals. We believe this method can provide a very useful tool for transferring knowledge from open loop to closed loop control on compliant robots.

Reservoir computing (RC) is a technique in machine learning inspired by neural systems. RC has been used successfully to solve complex problems such as signal classification and signal generation. These systems are mainly implemented in software, and thereby they are limited in speed and power efficiency. Several optical and optoelectronic implementations have been demonstrated, in which the system has signals with an amplitude and phase. It is proven that these enrich the dynamics of the system, which is beneficial for the performance. In this paper, we introduce a novel optical architecture based on nanophotonic crystal cavities. This allows us to integrate many neurons on one chip, which, compared with other photonic solutions, closest resembles a classical neural network. Furthermore, the components are passive, which simplifies the design and reduces the power consumption. To assess the performance of this network, we train a photonic network to generate periodic patterns, using an alternative online learning rule called first-order reduced and corrected error. For this, we first train a classical hyperbolic tangent reservoir, but then we vary some of the properties to incorporate typical aspects of a photonics reservoir, such as the use of continuous-time versus discrete-time signals and the use of complex-valued versus real-valued signals. Then, the nanophotonic reservoir is simulated and we explore the role of relevant parameters such as the topology, the phases between the resonators, the number of nodes that are biased and the delay between the resonators. It is important that these parameters are chosen such that no strong self-oscillations occur. Finally, our results show that for a signal generation task a complex-valued, continuous-time nanophotonic reservoir outperforms a classical (i.e., discrete-time, real-valued) leaky hyperbolic tangent reservoir (normalized root-mean-square errors=0.030 versus NRMSE=0.127).

Using data retrieved from the Puppy II robot at the University of Zurich (UZH), we show that machine learning techniques with non-linearities and fading memory are effective for terrain classification, both supervised and unsupervised, even with a limited selection of input sensors. The results indicate that most information for terrain classification is found in the combination of tactile sensors and proprioceptive joint angle sensors. The classification error is small enough to have a robot adapt the gait to the terrain and hence move more robustly.

In the field of Reservoir Computing, scaling the spectral radius of the weight matrix of a random recurrent neural network to below unity is a commonly used method to ensure the Echo State Property. Recently it has been shown that this condition is too weak. To overcome this problem, other - more involved - sufficient conditions for the Echo State Property have been proposed. In this paper we provide a large-scale experimental verification of the Echo State Property for large recurrent neural networks with zero input and zero bias. Our main conclusion is that the spectral radius method remains a valid indicator of the Echo State Property; the probability that the Echo State Property does not hold, drops for larger networks with spectral radius below unity, which are the ones of practical interest.

A model, predictor, or error estimator is often used by a feedback controller to control a plant. Creating such a model is difficult when the plant exhibits nonlinear behavior. In this paper, a novel online learning control framework is proposed that does not require explicit knowledge about the plant. This framework uses two learning modules, one for creating an inverse model, and the other for actually controlling the plant. Except for their inputs, they are identical. The inverse model learns by the exploration performed by the not yet fully trained controller, while the actual controller is based on the currently learned model. The proposed framework allows fast online learning of an accurate controller. The controller can be applied on a broad range of tasks with different dynamic characteristics. We validate this claim by applying our control framework on several control tasks: 1) the heating tank problem (slow nonlinear dynamics); 2) flight pitch control (slow linear dynamics); and 3) the balancing problem of a double inverted pendulum (fast linear and nonlinear dynamics). The results of these experiments show that fast learning and accurate control can be achieved. Furthermore, a comparison is made with some classical control approaches, and observations concerning convergence and stability are made.

Many applications of Reservoir Computing (and other signal processing techniques) have to deal with information processing of signals with multiple time-scales. Classical Reservoir Computing approaches can only cope with multiple frequencies to a limited degree. In this work we investigate reservoirs build of band-pass filter neurons which can be made sensitive to a specified frequency band. We demonstrate that many currently difficult tasks for reservoirs can be handled much better by a band-pass filter reservoir.