arXiv subjects

Chuhan Zhang

Publications and source records attributed to Chuhan Zhang.

At least 19 recordsLinked to original sources

arXiv2026

World-Coherent Decoding: Self-Verifying Test-Time Planning for World Action Models

World Action Models (WAMs) aim to control robots by stochastically generating visual futures and then decoding actions, but empirical observations indicate that the results can strongly depend on which future is selected. We propose World-Coherent-Decoding (WCD), a self-verifying test-time planning framework that treats WAM rollouts as falsifiable future--action hypotheses. At each decision step, WCD samples multiple candidates from a frozen WAM and ranks them using internal generative signals: flow-based video surprisal for visual plausibility and action path effort for action-generation stability. After execution, the realized observation audits the selected imagination, yielding an imagination--reality mismatch that trains a lightweight online predictor for future candidate selection. Thus, WCD converts delayed self-verification into pre-execution reliability estimation without updating the backbone model. On RoboTwin 2.0, WCD improves Hard success under limited randomized-scene supervision from $55.80\%$ to $60.90\%$, with a $+16.43$ gains on Horizon-3 tasks, and shows qualitative robustness on real Franka visual-shift tests. These results highlight a simple principle: test-time scaling for WAMs depends less on sampling more futures than on selecting reliable ones.

cs.CV↗

arXiv2026

Learning to infer and manipulate through distributed whole-arm interaction in a soft robot

In animals such as elephants and octopuses, acquiring non-visual information about an object and physically engaging with it are inseparable processes mediated by rich, large-area interactions between compliant appendages and the environment. Soft robots provide a natural platform for translating this principle into engineered systems. Yet current robotic intelligence makes limited use of physical interaction, treating it primarily as a disturbance to be rejected or, at best, as a means of compensating for object misalignment. Here, we introduce a physical intelligence framework in which distributed compliant interactions jointly reveal task-relevant information and organize manipulation behavior. This results in an intrinsically partially observable problem: key task-relevant information is never measured directly, but must instead be inferred from the history of physical interactions. We propose a reinforcement-learning architecture that addresses this challenge by learning a memory-based control policy end-to-end. The key innovations making this possible are (i) a pretrained exploration policy that provides a reference for broad workspace exploration, (ii) joint optimization that integrates exploration and grasping objectives within a single recurrent policy, and (iii) a two-stage sim-to-real adaptation including observation mapping and policy fine-tuning. We demonstrate this principle through blind whole-arm grasping with a hybrid rigid-soft robotic arm that we equip with IMUs embedded directly within its compliant structure, providing its only source of proprioceptive sensing. The learned policy successfully identifies and grasps various objects by autonomously coordinating workspace exploration, object encounter and localization, inference of grasp-relevant properties, and stable whole-arm wrapping.

cs.RO↗

arXiv2026

SoRoMoX: Fast, Differentiable, and Parallelizable Soft Robot Models

Reduced-order models based on Cosserat-rod theory are now well established, and modeling theory is no longer the primary bottleneck in soft-robot control. Their implementations, however, do not support the differentiable, GPU-parallel, and control-oriented workflows that underpin advanced rigid-robotics applications. Here, we fill this gap with SoRoMoX (Soft Robot Models in JAX), a fully numerical, JIT-compilable Python/JAX framework. SoRoMoX implements articulated, Piecewise Constant Strain, and Variable Strain models through a unified, control-ready interface that provides inertia matrices, gravitational and elastic forces, Jacobians, and their derivatives. To our knowledge, it is the first rod/strain-based soft-robot modeling framework that runs directly on GPUs and is end-to-end differentiable with respect to states, inputs, and parameters. Sequential CPU rollouts are up to 18.1x faster than state-of-the-art alternatives, while GPU-parallel rollouts increase throughput by up to 234.6x. This performance enables workflows that were previously impractical or impossible: static-equilibrium system identification with 66% lower marker RMSE; residual-force learning with a further 64% reduction; computed-torque tracking with RMSE reduced by a factor of approximately 500 relative to model-free PD; control-gain optimization with up to 62% lower loss than untuned gains; safety-constrained control using high-order control barrier functions to keep the peak contact force within a prescribed 5 N bound, compared with 33.5 N without the safety constraint; and reinforcement-learning policy training up to 7x faster than a CPU PyElastica discrete-rod baseline through massively parallel rollouts.

cs.RO↗

arXiv2026

Video Generation Models are General-Purpose Vision Learners

Driven by next-token prediction, NLP shifted from task-specific models into powerful generalist foundation models. What, then, is the equivalent catalyst needed to achieve a general-purpose model in computer vision? In this paper, we contend that large-scale text-to-video generation serves as a strong pre-training paradigm for computer vision, providing the necessary spatiotemporal priors, vision-language alignment, and scalability required for general visual intelligence. We introduce GenCeption, which leverages a pre-trained video generative diffusion backbone to define a feed-forward perception model, capable of performing various vision tasks steered by text instructions. Empirical results demonstrate that GenCeption achieves state-of-the-art performance across a diverse suite of tasks, including depth, surface normal, and camera pose estimation, expression-referring segmentation, and 3D keypoint prediction, often matching or surpassing specialized models (e.g. DepthAnything3, SAM3, D4RT, VGGT-Omega, Sapiens, David, Genmo, and Lotus-2). Furthermore, the video generative pretrained backbone outperforms alternative pretraining paradigms (e.g., V-JEPA, and Video MAE) under comparable settings. Importantly, GenCeption exhibits preliminary data and model scaling properties along with exceptional data efficiency, where it achieves comparable performance with leading models like D4RT and VGGT-Omega with 7 to 500 less training data. Finally, GenCeption also exhibits intriguing emergent behaviors: a model trained exclusively on synthetic human videos generalizes to real-world footage and out-of-distribution object categories (e.g., animals and robots). These findings suggest that video generation is not merely a synthesis tool, but a foundational path toward generalist vision intelligence for the physical world. Project page: https://genception.github.io

cs.CV↗

arXiv2026

Dynamics, stability, and energy efficiency of an energy-recycling rimless wheel with spring-clutch legs

This paper proposes an energy-recycling rimless wheel with spring-clutch legs. The proposed mechanism uses a lockable clutch to store part of the impact-induced elastic energy after foot contact and reinject it in the next gait cycle. First, we develop a hybrid dynamic model of the energy-recycling rimless wheel. Second, numerical simulations are used to examine the dynamics, local stability of periodic gaits, and the Cost of Transport (CoT) of the proposed mechanism. The simulation results show that the proposed mechanism reduces the CoT by up to 16.13% compared with a benchmark viscoelastic-legged rimless wheel with telescopic spring-damper legs. Compared with the rigid rimless wheel, the viscoelastic-legged and energy-recycling models reduce the CoT by more than 50%. The energy-recycling model also maintains locally stable periodic gaits over the tested slope and stiffness ranges. Finally, prototype experiments on an inclined plane are conducted to examine the feasibility of the proposed mechanism. The experimental results show that the proposed rimless wheel achieves passive walking on a shallow 1{\deg} slope, corresponding to a CoT of approximately 0.02. These results suggest that the proposed spring-clutch mechanism can improve the simulated walking efficiency of the energy-recycling rimless wheel, while the prototype experiments support the feasibility of passive walking with the mechanism.

cs.RO↗

arXiv2026

Propeller-Assisted Robust 3D Hopping Robot with Hierarchical Force Allocation

Monopedal hopping robots are conceptually simple but highly dynamic and inherently unstable. Achieving robust 3D hopping is still difficult because ground reaction forces are available only during the short stance phase, while the robot is underactuated in flight. A key unresolved issue is how to improve flight-phase control authority. Propeller assistance provides a promising solution, but it requires careful coordination of leg-generated contact forces and propeller thrusts across stance and flight. This paper presents Pro-OMEGA2, a propeller-assisted 3D monopedal hopping robot with an active 3-RSR parallel leg and a trunk-mounted tri-rotor for auxiliary attitude regulation. To address the force coordination challenge, we propose a Hierarchical Force Allocation (HFA) framework based on a single rigid body (SRB) model. The leg generates the main stance contact wrench, while the tri-rotor provides auxiliary attitude regulation, compensating the residual attitude moment in stance and maintaining attitude during flight. Real-robot experiments in indoor and outdoor scenarios demonstrate sustained 3D hopping, including terrain transitions and impulsive push recovery, validating robustness under unmodeled contact and external disturbances.

cs.RO↗

arXiv2026

Ripples of Stellar Enrichment (RoSE) - simulating element production and mixing in a Milky Way-mass galactic disc star-by-star

We present the Ripples of Stellar Enrichment (RoSE) simulations, which follow an isolated Milky Way-mass disc galaxy with star-by-star feedback and nucleosynthesis from five physical enrichment channels -- Wolf-Rayet stars, type II supernovae, type Ia supernovae, asymptotic giant branch stars, and neutron star mergers. We use these simulations to test how elements' diverse nucleosynthetic origins imprint spatial, temporal, and inter-element abundance correlations in gas and newly formed stars. We find that nucleosynthetic source composition is the primary organising principle of elemental structure: elements sharing a dominant production channel exhibit similar spatial and temporal statistics and their abundances are strongly correlated with one another, while mixed-source pairs are much more weakly correlated. We show that a simple linear regression model based only on how element pairs differ in their nucleosynthetic origin is able to predict, with high fidelity, how strongly their abundances correlate, in both interstellar medium gas and coeval stars. Comparison with Milky Way stellar abundance data shows encouraging qualitative agreement, with differences between simulations and observations comparable to the scatter between independent observational datasets. Together, these results show that the covariance of galactic abundances is largely predictable from the mixture of nucleosynthetic sources supplying each element: same-source elements remain tightly correlated in gas and young stars, whereas differing source mixtures produce systematically weaker correlations.

astro-ph.GA↗

arXiv2025

Efficiently Reconstructing Dynamic Scenes One D4RT at a Time

Understanding and reconstructing the complex geometry and motion of dynamic scenes from video remains a formidable challenge in computer vision. This paper introduces D4RT, a simple yet powerful feedforward model designed to efficiently solve this task. D4RT utilizes a unified transformer architecture to jointly infer depth, spatio-temporal correspondence, and full camera parameters from a single video. Its core innovation is a novel querying mechanism that sidesteps the heavy computation of dense, per-frame decoding and the complexity of managing multiple, task-specific decoders. Our decoding interface allows the model to independently and flexibly probe the 3D position of any point in space and time. The result is a lightweight and highly scalable method that enables remarkably efficient training and inference. We demonstrate that our approach sets a new state of the art, outperforming previous methods across a wide spectrum of 4D reconstruction tasks. We refer to the project webpage for animated results: https://d4rt-paper.github.io/.

cs.CV↗

arXiv2025

Metallicity fluctuation statistics in the interstellar medium and young stars -- II. Elemental cross-correlations and the structure of chemical abundance space

Observations of Milky Way stars by multiplexed spectroscopic instruments and of gas in nearby galaxies using integral field units have made it possible to measure the abundances of multiple elements in both the interstellar medium and the stars that form out of it. These observations have revealed complex correlations between elemental abundances, but thus far there has been no analytic theoretical framework to interpret these data. In this paper we extend the simple stochastically-forced diffusion model of Krumholz & Ting (2018), which has proven successful at explaining the spatial abundance patterns of single elements, to multiple elements, clarifying why elements are correlated and what controls their degree of correlation, and making quantitative predictions for the degree of correlation in both gas and young stars. We show that our results are qualitatively consistent with observed patterns, and point out how application of this theory to measured correlations should enable determination of currently unknown parameters describing r-process nucleosynthesis.

astro-ph.GA↗

arXiv2025

Evolution of action-space coherence in a Milky Way-like simulation

Efforts to dynamically trace stars back to the now-dissolved clusters in which they formed rely implicitly on the assumption that stellar orbital actions are conserved. While this holds in a static, axisymmetric potential, it is unknown how strongly the time-varying, non-axisymmetric structure of a real galactic disk drives action drift that inhibits cluster reconstruction. We answer this question using a high-resolution magnetohydrodynamic simulation of a Milky Way-like spiral disc galaxy. We show that, while stars experience significant action evolution over $\lesssim 100$ Myr, they do so in a correlated fashion whereby stars born in close proximity maintain very similar actions for up to 0.5 Gyr. The degree of coherence shows no significant dependence on galactocentric radius, but varies between action components: vertical actions decohere for stars born more than a few hundred parsecs apart (likely due to giant molecular clouds), while radial and azimuthal actions remain correlated on kiloparsec scales (likely influenced by spiral arms). We use our measurements of the rate of action decoherence to develop a probabilistic framework that lets us infer the initial sizes of the star cluster progenitors of present-day stellar streams from their measured action distributions, which we apply to 438 known moving groups. Our results suggest that most of these streams likely originated from compact clusters, but that a significant minority are instead likely to be resonant or dynamically induced structures. This method of classifying streams complements existing methods, optimises the use of expensive spectroscopic abundance measurements, and will be enhanced by the more precise kinematic data that will soon become available from \textit{Gaia} DR4.

astro-ph.GA↗

arXiv2025

Interaction-Centric Knowledge Infusion and Transfer for Open-Vocabulary Scene Graph Generation

Open-vocabulary scene graph generation (OVSGG) extends traditional SGG by recognizing novel objects and relationships beyond predefined categories, leveraging the knowledge from pre-trained large-scale models. Existing OVSGG methods always adopt a two-stage pipeline: 1) \textit{Infusing knowledge} into large-scale models via pre-training on large datasets; 2) \textit{Transferring knowledge} from pre-trained models with fully annotated scene graphs during supervised fine-tuning. However, due to a lack of explicit interaction modeling, these methods struggle to distinguish between interacting and non-interacting instances of the same object category. This limitation induces critical issues in both stages of OVSGG: it generates noisy pseudo-supervision from mismatched objects during knowledge infusion, and causes ambiguous query matching during knowledge transfer. To this end, in this paper, we propose an inter\textbf{AC}tion-\textbf{C}entric end-to-end OVSGG framework (\textbf{ACC}) in an interaction-driven paradigm to minimize these mismatches. For \textit{interaction-centric knowledge infusion}, ACC employs a bidirectional interaction prompt for robust pseudo-supervision generation to enhance the model's interaction knowledge. For \textit{interaction-centric knowledge transfer}, ACC first adopts interaction-guided query selection that prioritizes pairing interacting objects to reduce interference from non-interacting ones. Then, it integrates interaction-consistent knowledge distillation to bolster robustness by pushing relational foreground away from the background while retaining general knowledge. Extensive experimental results on three benchmarks show that ACC achieves state-of-the-art performance, demonstrating the potential of interaction-centric paradigms for real-world applications.

cs.CV↗

arXiv2025

ReSpark: Leveraging Previous Data Reports as References to Generate New Reports with LLMs

Creating data reports is a labor-intensive task involving iterative data exploration, insight extraction, and narrative construction. A key challenge lies in composing the analysis logic-from defining objectives and transforming data to identifying and communicating insights. Manually crafting this logic can be cognitively demanding. While experienced analysts often reuse scripts from past projects, finding a perfect match for a new dataset is rare. Even when similar analyses are available online, they usually share only results or visualizations, not the underlying code, making reuse difficult. To address this, we present ReSpark, a system that leverages large language models (LLMs) to reverse-engineer analysis logic from existing reports and adapt it to new datasets. By generating draft analysis steps, ReSpark provides a warm start for users. It also supports interactive refinement, allowing users to inspect intermediate outputs, insert objectives, and revise content. We evaluate ReSpark through comparative and user studies, demonstrating its effectiveness in lowering the barrier to generating data reports without relying on existing analysis code.

cs.HC↗

arXiv2025

Fundamental limits to orbit reconstruction due to non-conservation of stellar actions in a Milky Way-like simulation

The conservation of stellar actions is a fundamental assumption in orbit reconstruction studies in the Milky Way. However, the disc is highly dynamic, with time-dependent, non-axisymmetric features like transient spiral arms and giant molecular clouds (GMCs) driving local fluctuations in the gravitational potential on top of the near-axisymmetric background. Using high-resolution magnetohydrodynamic simulations that incorporate gas dynamics and star formation, we quantify the rate at which these effects drive non-conservation of the actions of young stars from Myr to Gyr timescales. We find that action evolution is well described as a logarithmic random walk, with vertical action evolving more rapidly than radial action; the diffusion rate associated with this random walk is weakly dependent on the stellar birth environment and scales approximately linearly with the galactic orbital frequency at a star's position. The diffusion rates we measure imply a fundamental limit of $\sim 100$ Myr as the timescale over which stellar orbits can be reliably reconstructed using methods that assume action conservation. By comparing diffusion rates for younger stars to those measured for an older and more vertically-extended control population, we conclude that radial action evolution is driven primarily by transient spiral arms, while vertical action evolution is driven by gravitational scattering off gaseous structures. Our results have significant implications for galactic archaeology and disc dynamics studies, necessitating a closer look at the timescales over which actions are assumed to be conserved in the disc.

astro-ph.GA↗

arXiv2025

NeuroBreak: Unveil Internal Jailbreak Mechanisms in Large Language Models

Jailbreak attacks bypass the safety alignment of large language models (LLMs) to elicit harmful outputs, yet the vast parameter space makes diagnosing the underlying failure mechanisms extremely challenging. We present NeuroBreak, a visual analytics system that helps experts progressively unpack jailbreak mechanisms from layer-level semantics down to neuron-level behaviors. A layer-wise probing pipeline traces how harmful representations evolve across layers, while a dual-dimensional character--behavior categorization reveals each safety-related neuron's inherent tendency and contextual contribution. These analyses are made interpretable through tailored visualization designs: a task-driven probing projection that reveals safety decision boundaries, a dual-stream semantic evolution flow that traces cross-layer semantic shifts, and a character--behavior chord graph that unifies neuron roles, attribution scores, and collaborative relations in a single view with in-situ causal verification. Quantitative evaluations and case studies show that NeuroBreak uncovers safety failure causes and provides actionable insights for strengthening LLM defenses.

cs.CR↗

arXiv2025

Scaling 4D Representations

Scaling has not yet been convincingly demonstrated for pure self-supervised learning from video. However, prior work has focused evaluations on semantic-related tasks $\unicode{x2013}$ action classification, ImageNet classification, etc. In this paper we focus on evaluating self-supervised learning on non-semantic vision tasks that are more spatial (3D) and temporal (+1D = 4D), such as camera pose estimation, point and object tracking, and depth estimation. We show that by learning from very large video datasets, masked auto-encoding (MAE) with transformer video models actually scales, consistently improving performance on these 4D tasks, as model size increases from 20M all the way to the largest by far reported self-supervised video model $\unicode{x2013}$ 22B parameters. Rigorous apples-to-apples comparison with many recent image and video models demonstrates the benefits of scaling 4D representations. Pretrained models are available at https://github.com/google-deepmind/representations4d .

cs.CV↗

arXiv2025

Cyclic Contrastive Knowledge Transfer for Open-Vocabulary Object Detection

In pursuit of detecting unstinted objects that extend beyond predefined categories, prior arts of open-vocabulary object detection (OVD) typically resort to pretrained vision-language models (VLMs) for base-to-novel category generalization. However, to mitigate the misalignment between upstream image-text pretraining and downstream region-level perception, additional supervisions are indispensable, eg, image-text pairs or pseudo annotations generated via self-training strategies. In this work, we propose CCKT-Det trained without any extra supervision. The proposed framework constructs a cyclic and dynamic knowledge transfer from language queries and visual region features extracted from VLMs, which forces the detector to closely align with the visual-semantic space of VLMs. Specifically, 1) we prefilter and inject semantic priors to guide the learning of queries, and 2) introduce a regional contrastive loss to improve the awareness of queries on novel objects. CCKT-Det can consistently improve performance as the scale of VLMs increases, all while requiring the detector at a moderate level of computation overhead. Comprehensive experimental results demonstrate that our method achieves performance gain of +2.9% and +10.2% AP50 over previous state-of-the-arts on the challenging COCO benchmark, both without and with a stronger teacher model.

cs.CV↗

arXiv2025

Learning from Streaming Video with Orthogonal Gradients

We address the challenge of representation learning from a continuous stream of video as input, in a self-supervised manner. This differs from the standard approaches to video learning where videos are chopped and shuffled during training in order to create a non-redundant batch that satisfies the independently and identically distributed (IID) sample assumption expected by conventional training paradigms. When videos are only available as a continuous stream of input, the IID assumption is evidently broken, leading to poor performance. We demonstrate the drop in performance when moving from shuffled to sequential learning on three tasks: the one-video representation learning method DoRA, standard VideoMAE on multi-video datasets, and the task of future video prediction. To address this drop, we propose a geometric modification to standard optimizers, to decorrelate batches by utilising orthogonal gradients during training. The proposed modification can be applied to any optimizer -- we demonstrate it with Stochastic Gradient Descent (SGD) and AdamW. Our proposed orthogonal optimizer allows models trained from streaming videos to alleviate the drop in representation learning performance, as evaluated on downstream tasks. On three scenarios (DoRA, VideoMAE, future prediction), we show our orthogonal optimizer outperforms the strong AdamW in all three scenarios.

cs.CV↗

arXiv2025

From Image to Video: An Empirical Study of Diffusion Representations

Diffusion models have revolutionized generative modeling, enabling unprecedented realism in image and video synthesis. This success has sparked interest in leveraging their representations for visual understanding tasks. While recent works have explored this potential for image generation, the visual understanding capabilities of video diffusion models remain largely uncharted. To address this gap, we systematically compare the same model architecture trained for video versus image generation, analyzing the performance of their latent representations on various downstream tasks including image classification, action recognition, depth estimation, and tracking. Results show that video diffusion models consistently outperform their image counterparts, though we find a striking range in the extent of this superiority. We further analyze features extracted from different layers and with varying noise levels, as well as the effect of model size and training budget on representation and generation quality. This work marks the first direct comparison of video and image diffusion objectives for visual understanding, offering insights into the role of temporal information in representation learning.

cs.CV↗