arXiv subjects

Sunwoo Park

Publications and source records attributed to Sunwoo Park.

10 recordsLinked to original sources

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

CRePE: Curved Ray Expectation Positional Encoding for Unified-Camera-Controlled Video Generation

Video world models should predict future appearance in a way that remains consistent with 3D scene structure, camera motion, and lens geometry. Existing attention-level camera encodings, however, either describe each token only by its viewing ray---without locating scene content along that ray---or assume pinhole projection, limiting camera control under wide-angle and fisheye lenses. We introduce Curved Ray Expectation Positional Encoding (CRePE), which represents each image token as a depth-aware distribution along its Unified Camera Model (UCM) ray and integrates the expected rotary positional phasor along the curved path this distribution traces when projected into each query view. CRePE is realized through a lightweight Geometric Attention Adapter on a frozen video diffusion transformer, with pseudo radial-distance supervision from a monocular geometry foundation model serving as a stabilizing anchor rather than an inference-time input. CRePE improves camera-control, lens, and orientation fidelity across pinhole, wide-angle, and fisheye settings, and transfers zero-shot to unseen real fisheye and diverse pinhole videos. Through Radial MixForcing, the same positional pathway further accepts externally supplied radial maps, enabling scene-geometry-conditioned generation and source-video motion transfer that follow the supplied geometry more faithfully than dedicated depth-conditioned baselines. CRePE thus offers a compact interface that unifies camera control, implicit 3D scene state, and external geometry control for video world models.

cs.CV↗

arXiv2026

Multi-THuMBS: Multi-person Tracking of 3D Human Meshes Beyond Video Shots

Tracking multi-person 3D human meshes from in-the-wild videos is a highly challenging problem due to complex interactions, frequent occlusions, and severe truncation inherent in unconstrained environments. While recent approaches have improved robustness against these issues, they largely overlook the critical challenge prevalent in real-world footage: frequent shot changes. These abrupt transitions in camera viewpoints often cause existing methods to lose track of human identities and fail in reconstructing temporally coherent trajectories. Although several recent works have explored 3D human mesh tracking under shot changes, they are still limited to single-person scenarios, making them inadequate for real-world videos where multiple people interact and appear simultaneously. To address this limitation, we propose Multi-THuMBS (Multi-person Tracking of 3D Human Meshes Beyond Video Shots) that leverages a state-of-the-art 3D scene prior to reconstruct the two boundary frames in a single shared 3D space. Human meshes are then registered within the shared 3D space, maintaining per-person identity and motion consistency across shot changes. Extensive experiments demonstrate that our approach yields significant improvements in 3D human mesh recovery, camera pose estimation, and identity tracking, thereby ensuring high-fidelity motion reconstruction with consistent identity preservation across shots compared to previous state-of-the-art methods.

cs.CV↗

arXiv2026

Geometric 4D Stitching for Grounded 4D Generation

Recent 4D generation methods complete scene-level missing information using generative models and reconstruct the scene into radiance-based representations. However, these pipelines often present geometric inconsistencies in the generated content, and the radiance-based reconstruction requires expensive optimization. Furthermore, radiance-based representations often absorb these geometric inconsistencies into their view-dependent nature, failing to enforce the grounded geometric consistency. To address these issues, we propose Geometric 4D Stitching, an efficient framework that explicitly identifies missing geometric regions and complements them with geometrically grounded 4D stitches. As a result, our method constructs 4D scene representations in under 10 minutes on a single NVIDIA RTX 5090 GPU per one-step scene expansion, while improving geometric consistency. Moreover, we demonstrate that our explicit 4D stitching supports interative expansion of 4D mesh as well as 4D scene editing.

cs.CV↗

arXiv2026

Evolution of UE in Massive MIMO Systems for 6G: From Passive to Active

As wireless networks continue to evolve, stringent latency and reliability requirements and highly dynamic channels expose fundamental limitations of gNB-centric massive multiple-input multiple-output (mMIMO) architectures, motivating a rethinking of the user equipment (UE) role. In response, the UE is transitioning from a passive transceiver into an active entity that directly contributes to system-level performance. In this context, this article examines the evolving role of the UE in mMIMO systems during the transition from fifth-generation (5G) to sixth-generation (6G), bridging third generation partnership project (3GPP) standardization, device implementation, and architectural innovation. Through a chronological review of 3GPP Releases 15 to 19, we highlight the progression of UE functionalities from basic channel state information (CSI) reporting to artificial intelligence (AI) and machine learning (ML)-based CSI enhancement and UE-initiated beam management. We further examine key implementation challenges, including multi-panel UE (MPUE) architectures, on-device intelligent processing, and energy-efficient operation, and then discuss corresponding architectural innovations under practical constraints. Using digital-twin-based evaluations, we validate the impact of emerging UE-centric functionalities, illustrating that UE-initiated beam reporting improves throughput in realistic mobility scenarios, while a multi-panel architecture enhances link robustness compared with a single-panel UE.

cs.IT↗

arXiv2025

Delay Time Characterization on FPGA: A Low Nonlinearity, Picosecond Resolution Time-to-Digital Converter on 16-nm FPGA using Bin Sequence Calibration

We present a Time-to-Digital Converter (TDC) implemented on a 16 nm Xilinx UltraScale Plus FPGA that achieves a resolution of 1.15 ps, RMS precision of 3.38 ps, a differential nonlinearity (DNL) of [-0.43, 0.24] LSB, and an integral nonlinearity (INL) of [-2.67, 0.15] LSB. This work introduces two novel hardware-independent post-processing techniques - Partial Order Reconstruction (POR) and Iterative Time-bin Interleaving (ITI) - that significantly enhance the performance of FPGA-based TDCs. POR addresses the missing code problem by inferring the partial order of each time bin through code density test data and directed acyclic graph (DAG) analysis, enabling near-complete recovery of usable bins. ITI further improves fine time resolution by merging multiple calibrated tapped delay lines (TDLs) into a single unified delay chain, achieving scalable resolution without resorting to averaging. Compared to state-of-the-art FPGA-based TDC architectures, the proposed methods deliver competitive or superior performance with reduced hardware overhead. These techniques are broadly applicable to high-resolution time measurement and precise delay calibration in programmable logic platforms.

cs.AR↗