arXiv subjects

Jangho Park

Publications and source records attributed to Jangho Park.

14 recordsLinked to original sources

arXiv2026

Foveated Compression: Selective High-Resolution Preservation for Token-Efficient VLMs

Visual tokens are a major source of inference cost in vision-language models, yet simple image downsampling remains a surprisingly strong compression baseline. This raises a complementary question: under a fixed token budget, where should visual fidelity be preserved? We introduce Foveated Compression, which encodes a full-resolution image once and represents it with a mixture of native- and compressed-resolution visual tokens. A behaviorally self-distilled Foveated Merger compresses local visual tokens while preserving compatibility with their native counterparts, and a lightweight Foveated Selector chooses one of nine spatial cells to retain at native resolution using exhaustive budget-matched intervention supervision. At 11.11% visual tokens, uniform Foveated Compression shows no significant paired difference from iso-token downsampling. At 20.99%, the learned selector significantly outperforms random and fixed allocation, but remains below strong whole-image resizing, showing that localized fidelity is not universally preferable. A budget-matched region-choice oracle reaches 82.73 macro accuracy versus 69.61 for the learned selector, revealing substantial headroom within the same spatial action space. Matched probing further shows that signals predicting when compression breaks the answer are substantially more accessible after language-model computation than to the lightweight prefill-free selector. These results expose complementary bottlenecks in region selection and compressed-region fidelity.

cs.CV↗

arXiv2026

FastOPD: On-Policy Distillation for Lightweight VLA Deployment

Vision-Language-Action (VLA) foundation models have scaled rapidly to enhance manipulation performance and generalizability, but this scaling incurs high computational costs that render real-world deployment increasingly challenging. Existing approaches typically mitigate this issue by designing smaller architectures or reducing the iterative denoising steps in flow-based policies. In this work, we propose FastOPD, a foundation-to-lightweight VLA framework that enables the practical deployment of large-scale VLAs through efficient on-policy distillation. Specifically, FastOPD adapts a flow map for single-state teacher supervision and combines it with a self-consistency objective to construct a compact student that learns the teacher dynamics. Furthermore, we theoretically demonstrate that minimizing this objective allows the distilled student to recover a distribution on par with that induced by an ideal few-step teacher model. We evaluate FastOPD across diverse foundation policies in simulation and real-world experiments. On LIBERO, FastOPD retains 84% of the performance of $π_{0.5}$ with only two inference steps, reducing inference latency by 78.1% while outperforming existing few-step distillation baselines in average success rate. With LingBot-VLA as the teacher, FastOPD improves the single-step success rate over the base student by 15.9 percentage points on RoboTwin 2.0. We further demonstrate its applicability to a World Action Model (WAM) and deploy a compact student distilled from MolmoAct2 on a real robot.

cs.RO↗

arXiv2026

Don't Throw Away the Tail: Action Upcycling for Policy Acceleration

Modern robot policies predict a chunk of future actions from a single observation, execute only a prefix, and discard the rest before replanning. Choosing the length of this prefix, the execution horizon, poses a trade-off between reactivity and efficiency. A short horizon keeps the policy reactive to the environment, but requires frequent policy calls. Recent test-time methods adaptively select the horizon for each chunk, but they either read model internals, where the signal must be chosen for each architecture, or draw extra samples, which adds cost. We propose Action Upcycling, a training-free algorithm that reuses actions the policy would otherwise discard, without accessing model internals or drawing extra samples. We find that discarded actions stay close to their replanned versions as long as the action velocity remains smooth. Action Upcycling therefore extends the execution horizon up to the point where the velocity begins to fluctuate. Extensive experiments on simulated and real-world manipulation tasks show that Action Upcycling reduces policy calls by 1.2-1.7x with no loss in success rate, across multiple Vision-Language-Action Models (VLAs) and even a World Action Model (WAM). It applies to any chunked policy at negligible cost and is orthogonal to other policy acceleration methods such as few-step sampling and streaming action decoding, opening a new axis for policy acceleration.

cs.RO↗

arXiv2026

DriftOPD: Sequence-Level Reverse-KL Distillation for One-Step VLA Policies

Vision-Language-Action (VLA) models increasingly rely on action experts that generate short action chunks under receding-horizon control. While chunk-level training is convenient across robot embodiments, it optimizes local action likelihood without explicitly accounting for long-horizon task success. Sequence-level reinforcement learning can address this limitation, but typically requires policy rollouts and closed-loop interaction, which are costly for real-robot manipulation. We introduce DriftOPD, a teacher-free, rollout-free framework for sequence-level on-policy distillation of continuous VLA action experts. We show that the sequence-level reverse Kullback-Leibler (KL) divergence decomposes into a chunk-level reverse-KL term and a future-potential term that captures the long-horizon effect of the current action. DriftOPD optimizes these two terms using a one-step drifting objective and a Q-function critic learned from offline demonstrations, respectively, enabling sequence-level optimization with only offline data and one-step action generation. Across multiple VLA architectures in simulation and real-world manipulation, DriftOPD generally outperforms existing one-step distillation baselines while achieving task success performance comparable to multi-step teacher policies. These results demonstrate that long-horizon behavior can be effectively distilled into one-step VLA action experts without online interaction or a separate teacher.

cs.RO↗

arXiv2026

Adjoint Guidance Flow: Amortized Critic Guidance for VLA Policies

Flow-based Vision-Language-Action (VLA) policies are typically trained by behavior cloning and thus do not explicitly optimize long-term task return. Critic guidance steers generation toward higher-value actions, but existing methods differentiate the critic through a one-step surrogate of the sampler and back-propagate a critic ensemble at every flow step. In contrast, here we propose Adjoint Guidance Flow (AGF), which amortizes trajectory-aware critic guidance into a lightweight guidance network while preserving the pretrained VLA policy. Specifically, we formulate critic-guided flow generation as a deterministic optimal control problem, whose optimal guidance is a costate that carries the terminal critic gradient back through the remaining flow, and regress the guidance network onto this costate while keeping both the VLA and critic frozen. This design provides favorable memory and throughput scaling during training, and inference needs one guidance-network forward pass per step, without the critic ensemble, back-propagation, or adjoint computation. Across LIBERO, RoboCasa, and LIBERO-Pro, AGF consistently improves pretrained VLAs, remains competitive with critic-guidance and policy-fine-tuning baselines, and is the most robust method when a single guidance strength is deployed across tasks. Compared with QGF, AGF runs $3.6\times$ faster per guidance step with $7.0\times$ fewer parameters, with comparable and even better performance, showing that critic guidance can be trajectory-aware and lightweight.

cs.LG↗

arXiv2026

Dynamic Manipulation with World-Action Models via Counterfactual Planning

World-Action models (WAMs) trained on static demonstrations often fail to manipulate moving targets even when they possess the required manipulation skills. We attribute this failure to target-response collapse: as execution advances, the policy becomes increasingly biased toward the learned continuation of its ongoing behavior and less responsive to target relocation. To bridge the gap between what the model has learned and what it can generate from the current context, we formulate dynamic manipulation as counterfactual planning by decoupling the context used for plan generation from the physical state used for execution. Our framework, Dynamic Predictive Planning (DPP), first uses the WAM's predictive rollout to estimate when an interaction is expected to occur, and combines this timing estimate with observed target motion to predict the target's future interaction position. DPP then constructs a counterfactual observation that places this predicted target position in a familiar robot context, allowing the model to invoke an existing manipulation skill rather than generate a recovery behavior from an unfamiliar robot-target configuration. The resulting plan is connected to the robot's actual state during execution. DPP enables real-time dynamic manipulation on a single consumer GPU without additional training on dynamic data. Experiments in simulation and on a real robot demonstrate consistent improvements across diverse target motions, with simulation performance surpassing all evaluated baselines, including methods additionally trained on dynamic data. Project page: https://methoder00.github.io/DPP/

cs.RO↗

arXiv2026

FlowLong: Inference-time Long Video Generation via Manifold-constrained Tweedie Matching

Extending the generation horizon of video diffusion models to long sequences remains a long-standing and important challenge. Existing training-free approaches fall into two categories: extensions of bidirectional models, which are tightly coupled to specific architectures and suffer from quality degradation over long horizons, and autoregressive models, which accumulate drift errors due to exposure bias and tend to produce repetitive motion patterns. To address these issues, we propose a novel but simple inference-time approach for long video generation that is architecture-agnostic and requires no additional training. Our method generates long videos via overlapping sliding windows, where predicted clean samples from adjacent windows are blended via \emph{Tweedie matching} to enforce both \textbf{manifold constraint and temporal consistency} across overlap regions. \emph{Stochastic early-phase sampling} then synchronizes per-window trajectories by injecting fresh noise after each Tweedie matching correction in the high-noise phase, before transitioning to deterministic ODE sampling to preserve fine-grained visual fidelity. Applied to various video generation models, our method generates videos several times longer than the native window length while outperforming both training-free and autoregressive baselines in temporal consistency and visual quality, and further extends to audio-video joint generation and text-to-3DGS without any fine-tuning.

cs.CV↗

arXiv2025

Zero4D: Training-Free 4D Video Generation From Single Video Using Off-the-Shelf Video Diffusion

Multi-view or 4D video generation has emerged as a significant research topic. Nonetheless, recent approaches to 4D generation still struggle with fundamental limitations, as they primarily rely on harnessing multiple video diffusion models with additional training or compute-intensive training of a full 4D diffusion model with limited real-world 4D data and large computational costs. To address these challenges, here we propose the first training-free 4D video generation method that leverages the off-the-shelf video diffusion models to generate multi-view videos from a single input video. Our approach consists of two key steps: (1) By designating the edge frames in the spatio-temporal sampling grid as key frames, we first synthesize them using a video diffusion model, leveraging a depth-based warping technique for guidance. This approach ensures structural consistency across the generated frames, preserving spatial and temporal coherence. (2) We then interpolate the remaining frames using a video diffusion model, constructing a fully populated and temporally coherent sampling grid while preserving spatial and temporal consistency. Through this approach, we extend a single video into a multi-view video along novel camera trajectories while maintaining spatio-temporal consistency. Our method is training-free and fully utilizes an off-the-shelf video diffusion model, offering a practical and effective solution for multi-view video generation.

cs.CV↗

arXiv2024

Unified Editing of Panorama, 3D Scenes, and Videos Through Disentangled Self-Attention Injection

While text-to-image models have achieved impressive capabilities in image generation and editing, their application across various modalities often necessitates training separate models. Inspired by existing method of single image editing with self attention injection and video editing with shared attention, we propose a novel unified editing framework that combines the strengths of both approaches by utilizing only a basic 2D image text-to-image (T2I) diffusion model. Specifically, we design a sampling method that facilitates editing consecutive images while maintaining semantic consistency utilizing shared self-attention features during both reference and consecutive image sampling processes. Experimental results confirm that our method enables editing across diverse modalities including 3D scenes, videos, and panorama images.

cs.CV↗

arXiv2024

ED-NeRF: Efficient Text-Guided Editing of 3D Scene with Latent Space NeRF

Recently, there has been a significant advancement in text-to-image diffusion models, leading to groundbreaking performance in 2D image generation. These advancements have been extended to 3D models, enabling the generation of novel 3D objects from textual descriptions. This has evolved into NeRF editing methods, which allow the manipulation of existing 3D objects through textual conditioning. However, existing NeRF editing techniques have faced limitations in their performance due to slow training speeds and the use of loss functions that do not adequately consider editing. To address this, here we present a novel 3D NeRF editing approach dubbed ED-NeRF by successfully embedding real-world scenes into the latent space of the latent diffusion model (LDM) through a unique refinement layer. This approach enables us to obtain a NeRF backbone that is not only faster but also more amenable to editing compared to traditional image space NeRF editing. Furthermore, we propose an improved loss function tailored for editing by migrating the delta denoising score (DDS) distillation loss, originally used in 2D image editing to the three-dimensional domain. This novel loss function surpasses the well-known score distillation sampling (SDS) loss in terms of suitability for editing purposes. Our experimental results demonstrate that ED-NeRF achieves faster editing speed while producing improved output quality compared to state-of-the-art 3D editing models.

cs.CV↗

arXiv2022

Long-Term Missing Value Imputation for Time Series Data Using Deep Neural Networks

We present an approach that uses a deep learning model, in particular, a MultiLayer Perceptron (MLP), for estimating the missing values of a variable in multivariate time series data. We focus on filling a long continuous gap (e.g., multiple months of missing daily observations) rather than on individual randomly missing observations. Our proposed gap filling algorithm uses an automated method for determining the optimal MLP model architecture, thus allowing for optimal prediction performance for the given time series. We tested our approach by filling gaps of various lengths (three months to three years) in three environmental datasets with different time series characteristics, namely daily groundwater levels, daily soil moisture, and hourly Net Ecosystem Exchange. We compared the accuracy of the gap-filled values obtained with our approach to the widely-used R-based time series gap filling methods ImputeTS and mtsdi. The results indicate that using an MLP for filling a large gap leads to better results, especially when the data behave nonlinearly. Thus, our approach enables the use of datasets that have a large gap in one variable, which is common in many long-term environmental monitoring observations.

cs.LG↗

arXiv2020

Variance Reduction for Sequential Sampling in Stochastic Programming

This paper investigates the variance reduction techniques Antithetic Variates (AV) and Latin Hypercube Sampling (LHS) when used for sequential sampling in stochastic programming and presents a comparative computational study. It shows conditions under which the sequential sampling with AV and LHS satisfy finite stopping guarantees and are asymptotically valid, discussing LHS in detail. It computationally compares their use in both the sequential and non-sequential settings through a collection of two-stage stochastic linear programs with different characteristics. The numerical results show that while both AV and LHS can be preferable to random sampling in either setting, LHS typically dominates in the non-sequential setting while performing well sequentially and AV gains some advantages in the sequential setting. These results imply that, given the ease of implementation of these variance reduction techniques, armed with the same theoretical properties and improved empirical performance relative to random sampling, AV and LHS sequential procedures present attractive alternatives in practice for a class of stochastic programs.

math.OC↗

arXiv2020

A Multistage Distributionally Robust Optimization Approach to Water Allocation under Climate Uncertainty

This paper investigates a Multistage Distributionally Robust Optimization (MDRO) approach to water allocation under climate uncertainty. The MDRO is formed by creating sets of conditional distributions (called conditional ambiguity sets) on a finite scenario tree. The distributions in the conditional ambiguity sets remain close to a nominal conditional distribution according a $ϕ$-divergence (e.g., Kullback-Liebler divergence, Hellinger distance, Burg entropy, etc.). The paper discusses a decomposition algorithm to solve the resulting MDRO and applies the modeling and solution techniques to allocate water in a rapidly-developing area of Tucson, Arizona. Tucson, like many arid and semi-arid regions around the world, faces considerable uncertainty in its ability to provide water for its citizens in the future. The primary sources of uncertainty in the Tucson region include (1) unpredictable population growth, (2) the availability of water from the Colorado River, and (3) the effects of climate variability on water consumption. This paper integrates forecasts for all these sources of uncertainty into a single optimization model for robust and sustainable water allocation. Then, it uses this model to analyze the value of constructing additional treatment facilities to reduce future water shortages. The results indicate that the MDRO approach can be very valuable for water managers by providing insights to minimize their risks and help them plan for the future.

math.OC↗

arXiv2020

Surrogate Optimization of Deep Neural Networks for Groundwater Predictions

Sustainable management of groundwater resources under changing climatic conditions require an application of reliable and accurate predictions of groundwater levels. Mechanistic multi-scale, multi-physics simulation models are often too hard to use for this purpose, especially for groundwater managers who do not have access to the complex compute resources and data. Therefore, we analyzed the applicability and performance of four modern deep learning computational models for predictions of groundwater levels. We compare three methods for optimizing the models' hyperparameters, including two surrogate model-based algorithms and a random sampling method. The models were tested using predictions of the groundwater level in Butte County, California, USA, taking into account the temporal variability of streamflow, precipitation, and ambient temperature. Our numerical study shows that the optimization of the hyperparameters can lead to reasonably accurate performance of all models (root mean squared errors of groundwater predictions of 2 meters or less), but the ''simplest'' network, namely a multilayer perceptron (MLP) performs overall better for learning and predicting groundwater data than the more advanced long short-term memory or convolutional neural networks in terms of prediction accuracy and time-to-solution, making the MLP a suitable candidate for groundwater prediction.

stat.ML↗