Utilizing pre-training holds great promise in legged robotics to produce effective movements. Han et al. propose a hierarchical framework that reuses pre-trained knowledge across various levels of task and perception. The cover image shows their quadrupedal robot MAX, developed by Tencent Robotics X, which demonstrates lifelike agility and strategic game-playing abilities.

See Han et al.

Estimating position and movement relative to objects with vision is a challenging computational problem in robotics. Renner et al. propose a neuromorphic solution, aiming at low-power, brain-inspired machine vision for mobile robots. In this approach, event-based neuromorphic vision sensors convert luminance changes during movement into spikes, illustrated in the cover image, with colour representing time. A neural architecture then analyses scene structure and the sensor’s pose, building a working memory of the environment.

See Renner et al. and Renner et al.

Quantum computing promises to be a transformative technology, but there are several challenges in realizing quantum computing hardware. One is the generation of quantum circuits that perform desired operations. Denoising diffusion models excel at this task, providing a powerful and flexible method to create circuits in a variety of scenarios. Given a text prompt that describes a quantum operation, they rely on iteratively denoising an initially noisy canvas until the desired quantum circuit is reached.

See Fürrutter et al.

Diffusion models can be used to generate intricate and detailed particle paths in turbulent flows, reflecting the complex nature of fluid motion. By means of statistical analysis, Li et al. show that diffusion models can capture the full complexity of turbulent dynamics and generalize to extreme events.

See Li et al.

Understanding the genetic factors that underlie the normal variation in cardiac shape is of great interest. In this work, Bonazzola et al. apply unsupervised geometric deep learning to phenotype the left ventricle by using an MRI-derived three-dimensional mesh representation (as depicted on the cover). The authors show that this approach boosts genetic discovery and provides deeper insights into the genetic underpinnings of cardiac morphology.

See Bonazzola et al.

Predicting the structure of 3D biological binding complexes is a major challenge in structural biology. Qiao et al. report a diffusion model-based generative AI approach known as NeuralPLexer that enables the prediction of protein–ligand structures, including large-scale conformational changes of such structures after ligand binding, based on protein sequences and ligand molecular graphs. The methodology could advance the mechanistic understanding of biological pathways and aid the discovery of new therapeutic agents.

See Qiao et al.

Training a neural network, which involves optimizing its parameters to reduce a loss function, can be thought of as moving through a landscape with hills of high error and valleys of low error. In the cover image, the red line shows such a trajectory, moving along the gradient towards lower loss. In this issue, Ciceri et al. describe that in successfully learning classification tasks, this training trajectory does not follow a direct route. Instead, the path takes a detour, shown here in brighter red, in which the representation of the data separates in training before later rejoining.

See Ciceri et al.