Search

When training transformers on graph-structured data, incorporating information about the underlying topology is crucial for good performance. Topological masking, a type of relative position encoding, achieves this by upweighting or downweighting attention depending on the relationship between the query and keys in a graph. In this paper, we propose to parameterise topological masks as a learnable function of a weighted adjacency matrix -- a novel, flexible approach which incorporates a strong structural inductive bias. By approximating this mask with graph random features (for which we prove the first known concentration bounds), we show how this can be made fully compatible with linear attention, preserving $\mathcal{O}(N)$ time and space complexity with respect to the number of input tokens. The fastest previous alternative was $\mathcal{O}(N \log N)$ and only suitable for specific graphs. Our efficient masking algorithms provide strong performance gains for tasks on image and point cloud data, including with $>30$k nodes.

The majority of language model training builds on imitation learning. It covers pretraining, supervised fine-tuning, and affects the starting conditions for reinforcement learning from human feedback (RLHF). The simplicity and scalability of maximum likelihood estimation (MLE) for next token prediction led to its role as predominant paradigm. However, the broader field of imitation learning can more effectively utilize the sequential structure underlying autoregressive generation. We focus on investigating the inverse reinforcement learning (IRL) perspective to imitation, extracting rewards and directly optimizing sequences instead of individual token likelihoods and evaluate its benefits for fine-tuning large language models. We provide a new angle, reformulating inverse soft-Q-learning as a temporal difference regularized extension of MLE. This creates a principled connection between MLE and IRL and allows trading off added complexity with increased performance and diversity of generations in the supervised fine-tuning (SFT) setting. We find clear advantages for IRL-based imitation, in particular for retaining diversity while maximizing task performance, rendering IRL a strong alternative on fixed SFT datasets even without online data generation. Our analysis of IRL-extracted reward functions further indicates benefits for more robust reward functions via tighter integration of supervised and preference-based LLM post-training.

Supersymmetry is an algebraic property of a class of quantum Hamiltonians, a guiding principle used in the development of nuclear theory and a concept that may underpin physics beyond the Standard Model. In condensed matter, any symmetry between bosonic and fermionic entities is usually strongly broken, as for the spin and charge sectors of the $t$-$J$ model describing the dynamics of fractionalized electrons in one-dimensional materials. While the triplet excitations of a quantum spin ladder in an applied magnetic field provide a close analogue of the $t$-$J$ chain, the supersymmetric nature of this system has not been explored. Here we perform neutron spectroscopy on the spin-ladder compounds (C$_5$D$_{12}$N)$_2$CuBr$_4$ and (C$_5$D$_{12}$N)$_2$CuCl$_4$ over a range of applied fields and temperatures, and apply matrix-product-state (MPS) methods to the ladder and equivalent chain models, to analyse the full, momentum-resolved dynamics of a single charge excitation in a bath of fractionalized spins. We find a dramatic difference in the effects of thermal fluctuations, where strong and coherent shifts of spectral weight at the zone edge contrast with a strict pole at the zone centre, whose persistence at all temperatures constitutes an observable consequence of supersymmetry in a quantum spin ladder.

Achieving human-level speed and performance on real world tasks is a north star for the robotics research community. This work takes a step towards that goal and presents the first learned robot agent that reaches amateur human-level performance in competitive table tennis. Table tennis is a physically demanding sport which requires human players to undergo years of training to achieve an advanced level of proficiency. In this paper, we contribute (1) a hierarchical and modular policy architecture consisting of (i) low level controllers with their detailed skill descriptors which model the agent's capabilities and help to bridge the sim-to-real gap and (ii) a high level controller that chooses the low level skills, (2) techniques for enabling zero-shot sim-to-real including an iterative approach to defining the task distribution that is grounded in the real-world and defines an automatic curriculum, and (3) real time adaptation to unseen opponents. Policy performance was assessed through 29 robot vs. human matches of which the robot won 45% (13/29). All humans were unseen players and their skill level varied from beginner to tournament level. Whilst the robot lost all matches vs. the most advanced players it won 100% matches vs. beginners and 55% matches vs. intermediate players, demonstrating solidly amateur human-level performance. Videos of the matches can be viewed at https://sites.google.com/view/competitive-robot-table-tennis, Comment: v2, 29 pages, 19 main paper, 10 references + appendix, adding an additional 9 references

We introduce T-CREx, a novel model-agnostic method for local and global counterfactual explanation (CE), which summarises recourse options for both individuals and groups in the form of human-readable rules. It leverages tree-based surrogate models to learn the counterfactual rules, alongside 'metarules' denoting their regions of optimality, providing both a global analysis of model behaviour and diverse recourse options for users. Experiments indicate that T-CREx achieves superior aggregate performance over existing rule-based baselines on a range of CE desiderata, while being orders of magnitude faster to run., Comment: Accepted at ICML 2024

Visual relationship detection aims to identify objects and their relationships in images. Prior methods approach this task by adding separate relationship modules or decoders to existing object detection architectures. This separation increases complexity and hinders end-to-end training, which limits performance. We propose a simple and highly efficient decoder-free architecture for open-vocabulary visual relationship detection. Our model consists of a Transformer-based image encoder that represents objects as tokens and models their relationships implicitly. To extract relationship information, we introduce an attention mechanism that selects object pairs likely to form a relationship. We provide a single-stage recipe to train this model on a mixture of object and relationship detection data. Our approach achieves state-of-the-art relationship detection performance on Visual Genome and on the large-vocabulary GQA benchmark at real-time inference speeds. We provide ablations, real-world qualitative examples, and analyses of zero-shot performance., Comment: ECCV camera-ready; changed to graph-constrained results

Large language models (LLMs) have been shown to exhibit a wide range of capabilities, such as writing robot code from language commands -- enabling non-experts to direct robot behaviors, modify them based on feedback, or compose them to perform new tasks. However, these capabilities (driven by in-context learning) are limited to short-term interactions, where users' feedback remains relevant for only as long as it fits within the context size of the LLM, and can be forgotten over longer interactions. In this work, we investigate fine-tuning the robot code-writing LLMs, to remember their in-context interactions and improve their teachability i.e., how efficiently they adapt to human inputs (measured by average number of corrections before the user considers the task successful). Our key observation is that when human-robot interactions are viewed as a partially observable Markov decision process (in which human language inputs are observations, and robot code outputs are actions), then training an LLM to complete previous interactions is training a transition dynamics model -- that can be combined with classic robotics techniques such as model predictive control (MPC) to discover shorter paths to success. This gives rise to Language Model Predictive Control (LMPC), a framework that fine-tunes PaLM 2 to improve its teachability on 78 tasks across 5 robot embodiments -- improving non-expert teaching success rates of unseen tasks by 26.9% while reducing the average number of human corrections from 2.4 to 1.9. Experiments show that LMPC also produces strong meta-learners, improving the success rate of in-context learning new tasks on unseen robot embodiments and APIs by 31.5%. See videos, code, and demos at: https://robot-teaching.github.io/.

Large, high-capacity models trained on diverse datasets have shown remarkable successes on efficiently tackling downstream applications. In domains from NLP to Computer Vision, this has led to a consolidation of pretrained models, with general pretrained backbones serving as a starting point for many applications. Can such a consolidation happen in robotics? Conventionally, robotic learning methods train a separate model for every application, every robot, and even every environment. Can we instead train generalist X-robot policy that can be adapted efficiently to new robots, tasks, and environments? In this paper, we provide datasets in standardized data formats and models to make it possible to explore this possibility in the context of robotic manipulation, alongside experimental results that provide an example of effective X-robot policies. We assemble a dataset from 22 different robots collected through a collaboration between 21 institutions, demonstrating 527 skills (160266 tasks). We show that a high-capacity model trained on this data, which we call RT-X, exhibits positive transfer and improves the capabilities of multiple robots by leveraging experience from other platforms. More details can be found on the project website https://robotics-transformer-x.github.io., Comment: Project website: https://robotics-transformer-x.github.io

Anticipating the motion of all humans in dynamic environments such as homes and offices is critical to enable safe and effective robot navigation. Such spaces remain challenging as humans do not follow strict rules of motion and there are often multiple occluded entry points such as corners and doors that create opportunities for sudden encounters. In this work, we present a Transformer based architecture to predict human future trajectories in human-centric environments from input features including human positions, head orientations, and 3D skeletal keypoints from onboard in-the-wild sensory information. The resulting model captures the inherent uncertainty for future human trajectory prediction and achieves state-of-the-art performance on common prediction benchmarks and a human tracking dataset captured from a mobile robot adapted for the prediction task. Furthermore, we identify new agents with limited historical data as a major contributor to error and demonstrate the complementary nature of 3D skeletal poses in reducing prediction error in such challenging scenarios., Comment: Project page: https://human-scene-transformer.github.io/

Zero-shot reinforcement learning (RL) promises to provide agents that can perform any task in an environment after an offline, reward-free pre-training phase. Methods leveraging successor measures and successor features have shown strong performance in this setting, but require access to large heterogenous datasets for pre-training which cannot be expected for most real problems. Here, we explore how the performance of zero-shot RL methods degrades when trained on small homogeneous datasets, and propose fixes inspired by conservatism, a well-established feature of performant single-task offline RL algorithms. We evaluate our proposals across various datasets, domains and tasks, and show that conservative zero-shot RL algorithms outperform their non-conservative counterparts on low quality datasets, and perform no worse on high quality datasets. Somewhat surprisingly, our proposals also outperform baselines that get to see the task during training. Our code is available via https://enjeeneer.io/projects/zero-shot-rl/., Comment: Project page: https://enjeeneer.io/projects/zero-shot-rl/

We present a deep-dive into a real-world robotic learning system that, in previous work, was shown to be capable of hundreds of table tennis rallies with a human and has the ability to precisely return the ball to desired targets. This system puts together a highly optimized perception subsystem, a high-speed low-latency robot controller, a simulation paradigm that can prevent damage in the real world and also train policies for zero-shot transfer, and automated real world environment resets that enable autonomous training and evaluation on physical robots. We complement a complete system description, including numerous design decisions that are typically not widely disseminated, with a collection of studies that clarify the importance of mitigating various sources of latency, accounting for training and deployment distribution shifts, robustness of the perception system, sensitivity to policy hyper-parameters, and choice of action space. A video demonstrating the components of the system and details of experimental results can be found at https://youtu.be/uFcnWjB42I0., Comment: Published and presented at Robotics: Science and Systems (RSS2023)

Quantum spin liquids (QSLs) have become a key area of research in magnetism due to their remarkable properties, such as long-range entanglement, fractional excitations, pinch-point singularities, and topologically protected phenomena. In recent years, the search for QSLs has expanded into the three-dimensional world, where promising features have been found in materials that form pyrochlore and hyper-kagome lattices, despite the suppression of quantum fluctuations due to high dimensionality. One such material is the $S = 1$ K$_2$Ni$_2$(SO$_4$)$_3$ compound, which belongs to the langbeinite family consisting of two interconnected trillium lattices. Although magnetically ordered, K$_2$Ni$_2$(SO$_4$)$_3$ has been found to exhibit a highly dynamical and correlated state which can be driven into a pure quantum spin liquid under magnetic fields of only $B \simeq 4$~T. In this article, we combine inelastic neutron scattering measurements with pseudo-fermion functional renormalization group (PFFRG) and classical Monte Carlo (cMC) calculations to study the magnetic properties of K$_2$Ni$_2$(SO$_4$)$_3$, revealing a high level of agreement between the experiment and theory. We further reveal the origin of the dynamical state in K$_2$Ni$_2$(SO$_4$)$_3$ by studying a larger set of exchange parameters, uncovering an `island of liquidity' around a focal point given by a magnetic network composed of tetrahedra on a trillium lattice.

We present an architecture and a training recipe that adapts pre-trained open-world image models to localization in videos. Understanding the open visual world (without being constrained by fixed label spaces) is crucial for many real-world vision tasks. Contrastive pre-training on large image-text datasets has recently led to significant improvements for image-level tasks. For more structured tasks involving object localization applying pre-trained models is more challenging. This is particularly true for video tasks, where task-specific data is limited. We show successful transfer of open-world models by building on the OWL-ViT open-vocabulary detection model and adapting it to video by adding a transformer decoder. The decoder propagates object representations recurrently through time by using the output tokens for one frame as the object queries for the next. Our model is end-to-end trainable on video data and enjoys improved temporal consistency compared to tracking-by-detection baselines, while retaining the open-world capabilities of the backbone detector. We evaluate our model on the challenging TAO-OW benchmark and demonstrate that open-world capabilities, learned from large-scale image-text pre-training, can be transferred successfully to open-world localization across diverse videos., Comment: ICCV 2023

The Ising chain realizes the fundamental paradigm of spin fractionalization, where locally flipping a spin creates two domain walls (spinons) that can separate apart at no energy cost. In a quasi-one-dimensional system, the mean-field effects of the weak three-dimensional couplings confine the spinons into a Zeeman ladder of two-spinon bound states. Here, we experimentally tune the confinement potential between spinons in the quasi-one-dimensional Ising ferromagnet CoNb2O6 by means of an applied magnetic field with a large component along the Ising direction. Using high-resolution single crystal inelastic neutron scattering, we directly observe how the spectrum evolves from the limit of very weak confinement at low field (with many closely-spaced bound states with energies scaling as the field strength to the power 2/3) to very strong confinement at high field (where it consists of a magnon and a dispersive two-magnon bound state, with a linear field dependence). At intermediate fields, we explore how the higher-order bound states disappear from the spectrum as they move to higher energies and overlap with the two-particle continuum. By performing a global fit to the observed spectrum in zero field and high field applied along two orthogonal directions, combined with a quantitative parameterization of the interchain couplings, we propose a refined single chain and interchain Hamiltonian that quantitatively reproduces all observed dispersions and their field dependence., Comment: 21 pages, 14 figures. Accepted version after comments from referees

We address a benchmark task in agile robotics: catching objects thrown at high-speed. This is a challenging task that involves tracking, intercepting, and cradling a thrown object with access only to visual observations of the object and the proprioceptive state of the robot, all within a fraction of a second. We present the relative merits of two fundamentally different solution strategies: (i) Model Predictive Control using accelerated constrained trajectory optimization, and (ii) Reinforcement Learning using zeroth-order optimization. We provide insights into various performance trade-offs including sample efficiency, sim-to-real transfer, robustness to distribution shifts, and whole-body multimodality via extensive on-hardware experiments. We conclude with proposals on fusing "classical" and "learning-based" techniques for agile robot control. Videos of our experiments may be found at https://sites.google.com/view/agile-catching, Comment: L4DC 2023

We propose a method to capture the handling abilities of fast jet pilots in a software model via reinforcement learning (RL) from human preference feedback. We use pairwise preferences over simulated flight trajectories to learn an interpretable rule-based model called a reward tree, which enables the automated scoring of trajectories alongside an explanatory rationale. We train an RL agent to execute high-quality handling behaviour by using the reward tree as the objective, and thereby generate data for iterative preference collection and further refinement of both tree and agent. Experiments with synthetic preferences show reward trees to be competitive with uninterpretable neural network reward models on quantitative and qualitative evaluations., Comment: arXiv admin note: substantial text overlap with arXiv:2210.01007

The MARI direct geometry time-of-flight neutron spectrometer at ISIS has been upgraded with an $m=3$ supermirror guide and new detector electronics. This has resulted in a flux gain of ${\approx}6{\times}$ at $\lambda=1.8$ {\AA}, and improvements on discriminating electrical noise, allowing MARI to continue to deliver a high quality science program well into its fourth decade of life.

This paper explores a rapid, optimal smooth path-planning algorithm for robots (e.g., autonomous vehicles) in point cloud environments. Derivative maps such as dense point clouds, mesh maps, Octomaps, etc. are frequently used for path planning purposes. A bi-directional target-oriented point planning algorithm, directly using point clouds to compute the optimized and dynamically feasible trajectories, is presented in this paper. This approach searches for obstacle-free, low computational cost, smooth, and dynamically feasible paths by analyzing a point cloud of the target environment, using a modified bi-directional and RRT-connect-based path planning algorithm, with a k-d tree-based obstacle avoidance strategy and three-step optimization. This presented approach bypasses the common 3D map discretization, directly leveraging point cloud data and it can be separated into two parts: modified RRT-based algorithm core and the three-step optimization. Simulations on 8 2D maps with different configurations and characteristics are presented to show the efficiency and 2D performance of the proposed algorithm. Benchmark comparison and evaluation with other RRT-based algorithms like RRT, B-RRT, and RRT star are also shown in the paper. Finally, the proposed algorithm successfully achieved different levels of mission goals on three 3D point cloud maps with different densities. The whole simulation proves that not only can our algorithm achieves a better performance on 2D maps compared with other algorithms, but also it can handle different tasks(ground vehicles and UAV applications) on different 3D point cloud maps, which shows the high performance and robustness of the proposed algorithm. The algorithm is open-sourced at \url{https://github.com/zhz03/BTO-RRT}, Comment: 12 Pages, 16 figures, submitted to T-IV and in review

The rare-earth triangular-lattice chalcogenide is a great platform for exploring both spin liquids and novel magnetic orders with anisotropic spin interactions and magnetic frustrations. Here, we report the thermodynamic and neutron scattering measurements of rare-earth triangular-lattice chalcogenide KErSe$_{2}$, using single-crystal samples. Our experiments revealed a long-range stripe order below 0.2 K. Although the magnetic order was three-dimensional, magnetic excitations exhibited negligible modulation along the z direction, indicating very weak interlayer coupling. Furthermore, magnetic excitation developed a well-defined spin-wave dispersion with a gap of $\sim$0.03 meV at M points. Both the stripe order and spin-wave excitations could be quantitatively understood from the anisotropic spin interactions of the Er$^{3+}$ Kramers doublets.

Recent efforts to learn reward functions from human feedback have tended to use deep neural networks, whose lack of transparency hampers our ability to explain agent behaviour or verify alignment. We explore the merits of learning intrinsically interpretable tree models instead. We develop a recently proposed method for learning reward trees from preference labels, and show it to be broadly competitive with neural networks on challenging high-dimensional tasks, with good robustness to limited or corrupted data. Having found that reward tree learning can be done effectively in complex settings, we then consider why it should be used, demonstrating that the interpretable reward structure gives significant scope for traceability, verification and explanation., Comment: 22 pages (9 main body). Preprint, under review

The collision efficiency of uncharged micron-sized water droplets in air is determined by the breakdown of hydrodynamics at droplet separations of the order of the mean-free path, by van-der-Waals forces, or a combination of the two. In contrast, electrostatic forces determine the collision efficiency of charged droplets if the charge is large enough. To find the charge for which the transition to charge-dominated collisions occurs, we computed the collision efficiency of charged, hydrodynamically-interacting droplets settling in quiescent air, including the breakdown of hydrodynamics at small interfacial distances. For oppositely charged droplets, the transition occurs when a saddle point of the relative droplet-dynamics exits the region where the hydrodynamics breaks down. For droplets with radii $16\,\mu$m and $20\,\mu$m, this occurs at $\sim 10^3$ elementary charges $e$. For smaller charges, the collision efficiency depends upon the Kn number (defined as the ratio of the mean-free-path of air to the mean droplet radius), whereas for larger charges it does not. For droplets charged with the same polarity, the critical charge is $\sim 10^4\,e$ for the above radii., Comment: 14 pages, 2 figures

Sim-to-real transfer is a powerful paradigm for robotic reinforcement learning. The ability to train policies in simulation enables safe exploration and large-scale data collection quickly at low cost. However, prior works in sim-to-real transfer of robotic policies typically do not involve any human-robot interaction because accurately simulating human behavior is an open problem. In this work, our goal is to leverage the power of simulation to train robotic policies that are proficient at interacting with humans upon deployment. But there is a chicken and egg problem -- how to gather examples of a human interacting with a physical robot so as to model human behavior in simulation without already having a robot that is able to interact with a human? Our proposed method, Iterative-Sim-to-Real (i-S2R), attempts to address this. i-S2R bootstraps from a simple model of human behavior and alternates between training in simulation and deploying in the real world. In each iteration, both the human behavior model and the policy are refined. For all training we apply a new evolutionary search algorithm called Blackbox Gradient Sensing (BGS). We evaluate our method on a real world robotic table tennis setting, where the objective for the robot is to play cooperatively with a human player for as long as possible. Table tennis is a high-speed, dynamic task that requires the two players to react quickly to each other's moves, making for a challenging test bed for research on human-robot interaction. We present results on an industrial robotic arm that is able to cooperatively play table tennis with human players, achieving rallies of 22 successive hits on average and 150 at best. Further, for 80% of players, rally lengths are 70% to 175% longer compared to the sim-to-real plus fine-tuning (S2R+FT) baseline. For videos of our system in action, please see https://sites.google.com/view/is2r., Comment: 8+24 pages

We generalise the problem of reward modelling (RM) for reinforcement learning (RL) to handle non-Markovian rewards. Existing work assumes that human evaluators observe each step in a trajectory independently when providing feedback on agent behaviour. In this work, we remove this assumption, extending RM to capture temporal dependencies in human assessment of trajectories. We show how RM can be approached as a multiple instance learning (MIL) problem, where trajectories are treated as bags with return labels, and steps within the trajectories are instances with unseen reward labels. We go on to develop new MIL models that are able to capture the time dependencies in labelled trajectories. We demonstrate on a range of RL tasks that our novel MIL models can reconstruct reward functions to a high level of accuracy, and can be used to train high-performing agent policies., Comment: 27 pages (10 main content; 2 references; 1 checklist; 14 appendix). 14 figures (9 main content; 5 appendix). Published at NeurIPS 2022. Revisions: v2) Updated to NeurIPS camera ready version (extra experiments)

When spin-orbit-entangled $d$-electrons reside on a honeycomb lattice, rich quantum states are anticipated to emerge, as exemplified by the $d^5$ Kitaev materials. Distinct yet equally intriguing physics may be realized with a $d$-electron count other than $d^5$. We found that the layered ruthenate Ag$_3$LiRu$_2$O$_6$ with $d^4$ Ru$^{4+}$ ions at ambient pressure forms a honeycomb lattice of spin-orbit-entangled singlets, which is a playground for frustrated excitonic magnetism. Under pressure, the singlet state does not develop the expected excitonic magnetism but experiences two successive transitions to other nonmagnetic phases, first to an intermediate phase with moderate distortion of honeycomb lattice, and eventually to a high-pressure phase with very short Ru-Ru dimer bonds. While the strong dimerization in the high-pressure phase originates from a molecular orbital formation as in the sister compound Li$_2$RuO$_3$, the intermediate phase represents a spin-orbit-coupled $J$-dimer state which is stabilized by the admixture of upper-lying $J_{\rm eff} = 1$-derived states. We argue that the $J$-dimer state is induced by a pseudo-Jahn-Teller effect associated with the low-lying spin-orbital excited states and is unique to spin-orbit-entangled $d^4$ systems. The discovery of competing singlet phases demonstrates rich spin-orbital physics of $d^4$ honeycomb compounds and paves the way for realization of unconventional magnetism., Comment: 12 pages, 5 figures

This work addresses the problems of semantic segmentation and image super-resolution by jointly considering the performance of both in training a Generative Adversarial Network (GAN). We propose a novel architecture and domain-specific feature loss, allowing super-resolution to operate as a pre-processing step to increase the performance of downstream computer vision tasks, specifically semantic segmentation. We demonstrate this approach using Nearmap's aerial imagery dataset which covers hundreds of urban areas at 5-7 cm per pixel resolution. We show the proposed approach improves perceived image quality as well as quantitative segmentation accuracy across all prediction classes, yielding an average accuracy improvement of 11.8% and 108% at 4x and 32x super-resolution, compared with state-of-the art single-network methods. This work demonstrates that jointly considering image-based and task-specific losses can improve the performance of both, and advances the state-of-the-art in semantic-aware super-resolution of aerial imagery., Comment: 11 pages, 7 figures

Water droplets coalesce into larger ones in atmospheric clouds to form rain. But droplets on collision courses do not always coalesce due to the cushioning effects of the air between them. The extent to which these so-called hydrodynamic interactions reduce coalescence rates is embodied in the collision efficiency, which is often small and is not generally known. In order to characterize the mechanisms that determine the collision efficiency, we exploited new time-resolved three-dimensional droplet tracking techniques to measure the positions of cloud droplet pairs settling through quiescent air. We did so with an unprecedented precision that enabled us to calculate the relative positions, velocities, and accelerations of the droplets at droplet surface-to-surface separations as small as about one-tenth of a droplet diameter. We show that relative accelerations clearly distinguish coalescing from non-coalescing droplet trajectories, the former being associated with relative accelerations that exceeded a threshold value. We outline how relative accelerations relate to hydrodynamic interactions, and present scaling arguments that predict the threshold relative acceleration. We speculate that the relative acceleration distribution of droplets in turbulent clouds can parameterize the collision efficiency, and that this distribution together with the well-known relative position and velocity distributions can generate a physical description of both the collision and coalescence rates of cloud droplets., Comment: 12 pages, 5 figures

The motion of a spin excitation across topologically non-trivial magnetic order exhibits a deflection that is analogous to the effect of the Lorentz force on an electrically charged particle in an orbital magnetic field. We used polarized inelastic neutron scattering to investigate the propagation of magnons (i.e., bosonic collective spin excitations) in a lattice of skyrmion tubes in manganese silicide. For wave vectors perpendicular to the skyrmion tubes, the magnon spectra are consistent with the formation of finely spaced emergent Landau levels that are characteristic of the fictitious magnetic field used to account for the nontrivial topological winding of the skyrmion lattice. This provides evidence of a topological magnon band structure in reciprocal space, which is borne out of the nontrivial real-space topology of a magnetic order.